You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@ossie.apache.org by "Ashfaqbs (via GitHub)" <gi...@apache.org> on 2026/10/04 07:25:54 UTC

[D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

GitHub user Ashfaqbs created a discussion: Proposal: an MCP server for Ossie (and where it should live)

## Proposal

I'd like a maintainer steer before building anything: does an MCP server for Ossie (and, downstream of that, a thin agent built on it) fit this project, and if so, where should it live?

### What it would do

An MCP server exposing Ossie semantic models as tools an AI agent can call directly, instead of an agent reading/writing raw YAML by hand or guessing at the schema. A rough first cut:

- `validate_model` - validate a semantic model document against the Ossie schema (wraps `validation/`), returning structured errors an agent can act on instead of parsing CLI output.
- `list_vendors` / `describe_dialect` - surface the registered vendor names and `Dialect` enum values from `core-spec/spec.yaml` so an agent doesn't have to guess which converter or expression dialect exists.
- `convert_model` - invoke a given converter (dbt, Salesforce, Power BI, etc.) in either direction, returning the converted document plus any warnings the converter logged.
- `query_model` - answer structural questions about a loaded model (list datasets/metrics/relationships, find a metric's dependencies, find custom_extensions for a given vendor) without the agent having to walk the YAML itself.

### Why this seems worth asking about

Discussion #374 (downstream agent field reports) already notes that coding/autonomous agents are a real, heavy consumer of this spec, and that that usage produces its own friction (issue-sprawl from agents guessing at the schema). An MCP server is a fairly direct answer to "give agents a correct, narrow interface instead of letting them freelance against raw YAML" - it would reduce exactly the kind of fragmented, context-free issues that discussion is trying to get ahead of.

### Where should this live?

I saw #453 proposes that new converters go to separate repositories (`apache/ossie-converters`, or coordinate on the dev mailing list) rather than the main repo. An MCP server isn't a converter, but it's the same shape of question - new tooling built on top of the spec rather than the spec itself. I'd rather ask than assume: should this go in a new repo under the Ossie project (e.g. alongside `ossie-converters`), as a community project referenced from the README, or is this out of scope for the project entirely?

Happy to scope a first PR to just `validate_model` and `list_vendors`/`describe_dialect` against the Python validation tooling already in `validation/` if there's interest, once there's a steer on where it should live.

GitHub link: https://github.com/apache/ossie/discussions/510

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
Ontology vocab vs reactive filter values vs pro-active, some trade-offs


  *
Ontology vocab: If you context window size allows, for sure its better to just let the agent know what the values are
  *
Pre-active filter: one can also expose a dedicated tool to enable an agent to find column values. We've seen good results in some scenarios  - one needs to consider the total number of tools and the agent might still end up picking the wrong filter
  *
Reactive filter values: I don't see too much harm in providing it, one can tune the scenarios when one returns them (e.g. 0-row results)

Also I don't see all the above as either-or, depending on data size, agent harness, ... one should probably support all.
________________________________
From: Kyoung Min Kim <ka...@gmail.com>
Sent: Wednesday, October 7, 2026 9:02 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

A data point on Chris's question about the input shape, from an experiment I
ran this week on real tables.

I took one database of 33 datasets, most of them monthly profile marts, with
45 relationships, generated an Ossie model for it, and gave three models
(one current, two older or smaller) a catalog of the datasets, fields and
three named metrics. Their only job was to return a spec in the Layer 3
shape: a metric or an aggregate, dimensions, filters. A compiler resolved
the joins from the relationships and produced the SQL. I scored ten
questions by compiling the reference spec and the model's spec and comparing
the SQL, counting result-equivalent variants as correct, and then ran the
pairs that differed against the warehouse to check those equivalence calls.
The current model got 8 of 10, the other two 7 and 8.

What made the difference was not the models but what the compiler could
check because the input was tied to the Ossie model. When a model picked a
dataset with no relationship path to the others, compilation refused instead
of producing a cross join. A filter on the period column of one monthly
dataset could be carried to the joined one, so the one spec that dropped it
did not count across months. And where a dataset declared unique_keys,
COUNT(*) against COUNT(DISTINCT key) could be proven equivalent or not,
which the warehouse run confirmed in both directions. The keys also
explained a mistake of my own: an alternative answer I had accepted counted
distinct values of a column that is only part of the declared key, and on
real data it undercounted by more than ten times. All of that is much harder
to offer once the server receives an engine-native string.

So I would keep a Layer 3 style input as the portable path and treat the
native-query input in #529 as the escape hatch Chris describes, with the
model identity required rather than optional whenever the request claims to
be about an Ossie model.

One more thing that touches Markus's filter value alternatives. The most
common remaining mistake involved code values: inventing a label for a
job-segment column, or not knowing that a filter on a code column was needed
at all, because nothing in the model lists the admissible values. The
ontology spec already has a place for that in a ValueType's requires list.
If execute_query returns value alternatives after the fact, it is worth
deciding whether that vocabulary also belongs in the model, where the agent
sees it before it queries.

Kyoung Min Kim


> 2026. 10. 7. AM 8:51, Aleksandr Kosachev <a....@gmail.com> wrote:
>
> Hi Chris,
>
> Two things would make it easier:
>
> - The spec should require a core entry in the document, not only support for it in the reader.
> - Whatever changes a query result should not sit in custom_extensions, which the spec calls metadata and which a reader from another vendor cannot interpret.
>
> === Why the first ===
>
> I ran the question against what the converters produce on their own vendor-format inputs, plus the examples: current main, 19 documents, 411 field and metric expressions. 32 of them carry a vendor dialect only.
>
> The 21 tagged DATABRICKS are plain SQL: bare columns, SUM, COUNT(*), COUNT(DISTINCT ...). ossie-clickhouse parses them unchanged and still cannot use them, because nothing in the document says this text is also core, and the spec lets a reader ignore other dialects. Every expression in those documents is affected, so hiding them would leave nothing to serve. I could transpile DATABRICKS with SQLGlot, and for these 21 it would work. But then a reader relies on the vendor entry and the core entry becomes the optional one, the opposite of what the spec says.
>
> The other 11 are THOUGHTSPOT, and they show where the gap is. In the TPC-DS fixture every field carries ANSI_SQL next to its formula; none of the 10 metrics does. The fixture is small, but #462 measured the same thing on 31 real ThoughtSpot models: 476 of 476 metrics have no portable dialect, and the Snowflake and Databricks converters drop them on the way out.
>
> Reading the formula as SQL anyway is worse than refusing it. 9 of the 11 go through a SQL parser without an error, with [store_sales::ss_ext_sales_price] coming out as a CAST, so the mistake would show up in the numbers. "Other dialects MAY be ignored" is the right rule as long as there is a core entry to fall back on.
>
> What it costs:
>
> Databricks: for these 21, one more line with the same text.
> ThoughtSpot: the measure column declared as a field, and the metric written over it. #463 tried SUM(dataset.column) and withdrew it, since readers split on whether that names a field or a column. #486 settles that it names a field.
> If a required core entry is too much for fields, metrics alone would close the gap. A weaker form would be a machine-readable statement that an expression has no core form, so a compiler can hide one metric instead of refusing the document.
>
> === Why the second ===
>
> Of the 79 custom_extensions entries, most are display formats, connection names and round-trip witnesses. One kind is different: the Databricks fixtures keep a model-level filter there (ss_net_profit > 0), and the core has no place for it. A reader from another vendor ignores the extension, so on the same document it returns different numbers from Databricks, and nothing in the document warns about that.
>
> The multiplicity thread is moving relationship cardinality into core. A model filter deserves the same question. For the rest, the per-converter schema documentation you +1'd would at least let a compiler see what it is ignoring.
>
> Kind regards,
> Alex
>
>> On 6 Oct 2026, at 14.52, Chris Eubank via dev <de...@ossie.apache.org> wrote:
>>
>> Hi Alex,
>>
>> I think we'll want to support both query interfaces wherever possible. #452
>> brings up different cases where one layer is more natural. For agents, I'd
>> expect both would be useful: layer 3 for handling many common analyses, and
>> layer 2 for when the agent wants to perform analysis not directly supported
>> by the model (e.g. joining in another table, tweaking a relationship) while
>> still having some guardrails (e.g. keeping the measure abstraction vs.
>> falling back to raw SQL).
>>
>> It is worth noting a few limitations: Ossie currently lacks official query
>> semantics. #452 points to the layered PRs, but those are still in flight.
>> There are also some foundational data modeling PRs (#343, #486) that would
>> be great to land, since some core behavior (e.g. name resolution rules)
>> remains underspecified.
>>
>> I would be really interested in your perspective on how you think about
>> querying Ossie models, especially since you're building on top of the
>> project. In this thread
>> <https://lists.apache.org/thread/r1vhmw2vdms3bvv1s9tsfm06knnnbn30> we were
>> discussing the possibility of vendor-specific extensions (both in
>> custom_extensions and expressions) being made more transparent. The
>> expression spec only states that compliant implementations must support
>> Ossie SQL, but I suspect in practice many of these vendor extensions will
>> exist in models. What would make your job easier when building your query
>> compilation on top of real-world models?
>>
>> Kind regards,
>> Chris
>>
>>
>> On Mon, Oct 5, 2026 at 4:23 PM Aleksandr Kosachev <a....@gmail.com>
>> wrote:
>>
>>> Hi Markus, JB,
>>>
>>> One gap the thread has not named yet. The REST API spec proposed on 09-23
>>> covers two things: managing models and binding a model to physical
>>> entities. It does not cover running a query and returning data. That is
>>> where execute_query starts. So an MCP spec next to the REST spec would not
>>> duplicate it. It would fill the missing step.
>>>
>>> Markus asked how an agent that found a model knows how to run a query
>>> against the source.
>>> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
>>> does not need to know:
>>> - The model carries no connection.
>>> - The server starts with a model and a database connection.
>>> - The agent sees only the tools.
>>>
>>> This reconciles the two positions in the thread. Ossie specifies the tool
>>> contract. The engine or catalog that serves it owns the binding.
>>>
>>> One question needs an answer before a spec: what is the input of
>>> execute_query? PR #452 names two query interfaces:
>>> - Layer 2: SQL with MEASURE().
>>> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>>>
>>> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
>>> When its planner refuses a query, the refusal names the cause.
>>> For example, "unknown metric 'x' (did you mean: ...)", or which metric
>>> would be counted once per joined row.
>>> The agent reads that and repairs the query.
>>> Markus listed four outputs: data, errors, suggestions, filter value
>>> alternatives.
>>> ossie-clickhouse returns data, errors and name suggestions, plus the
>>> generated SQL. It has no filter value alternatives yet.
>>> I would like to hear how those look on your side.
>>>
>>> On query_model and the harness dependency, ossie-clickhouse confirms it.
>>> Its overview tool returns datasets and metrics without fields.
>>> Agents then call a search tool and a describe-one-object tool before every
>>> query. A harness with the file on disk, as Markus described, does not need
>>> them. So discovery tools belong in the spec as optional, not dropped.
>>>
>>> Warm regards,
>>> Alex Kosachev
>>>
>>> [1] https://github.com/Dactopus/ossie-clickhouse
>>> <
>>> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
>>>>
>>>
>>> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>>>
>>>> I'm curious about the expectations here. If e.g. GitHub copilot discovers
>>>> an Ossie semantic model, how would it know how to connect/execute a query
>>>> against the data source (e.g. PBI semantic model, Databricks, Snowflake,
>>>> ...)? Is the expectation that the harness would get additional
>>> instructions
>>>> anyway?
>>>>
>>>> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
>>>> implementation.
>>>> ________________________________
>>>> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
>>>> Sent: Monday, October 5, 2026 8:25 PM
>>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>>> live) [ossie]
>>>>
>>>> Hi
>>>>
>>>> It's not clear to me if we are discussing the specification or the
>>>> implementation. MCP implementation in Ossie: I'm not convinced; it seems
>>>> more like an engine/catalog implementation.
>>>> The spec could make sense.
>>>>
>>>> But do we really need the MCP spec in Ossie for agent model discovery?
>>> I'm
>>>> not sure. That's actually one of the reasons for the REST API Spec too
>>>> (model discovery, versioning, etc).
>>>> We could have the MCP spec alongside the REST API Spec.
>>>>
>>>> Also, MCP is good to have but clearly not a requirement for the agent
>>> (API
>>>> or CLI are classic data providers for agent, including operations).
>>>>
>>>> Regards
>>>> JB
>>>>
>>>> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>>>>
>>>>> I'd like to provide a slightly different pitch for an MCP server spec
>>>> (not
>>>>> necessarily the implementation). From my perspective, Ossies biggest
>>>> value
>>>>> prop is related to empowering agents to understand semantics and
>>>>> subsequently build queries to retrieve data.
>>>>> Most important is probably an execute_query MCP tool.
>>>>>
>>>>>
>>>>> *
>>>>> Input: data source id, query
>>>>> *
>>>>> Output
>>>>>    *
>>>>> Data
>>>>>    *
>>>>> Errors
>>>>>    *
>>>>> Suggestions
>>>>>    *
>>>>> Filter value alternatives
>>>>>    *
>>>>> ...
>>>>>
>>>>> This operates under the assumption that the agent is able to
>>> discover/see
>>>>> the Ossie semantic model (e.g. using
>>>> http://agenticresourcediscovery.org/).
>>>>> We already have a few variants of execute_query tool across our
>>> engines.
>>>>>
>>>>> I'm not sure about query_model (yet), at least for now we see a
>>>> dependency
>>>>> on the agent harness. Some harness can only digest a small view of the
>>>>> semantic model and thus need tools to progressively discover, others
>>>> (like
>>>>> e.g. GitHub copilot cli), don't need any of that, as they'll have full
>>>>> filesystem access, download/query local.
>>>>>
>>>>> ________________________________
>>>>> From: jbonofre (via GitHub) <gi...@apache.org>
>>>>> Sent: Monday, October 5, 2026 8:01 AM
>>>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>>>> live) [ossie]
>>>>>
>>>>>
>>>>> GitHub user jbonofre added a comment to the discussion: Proposal: an
>>> MCP
>>>>> server for Ossie (and where it should live)
>>>>>
>>>>> I'm not sure we need MCP server at all in Ossie. That could be
>>> something
>>>>> catalogs can do.
>>>>>
>>>>> I would rather not consider MCP service directly in Ossie for now (I
>>>> don't
>>>>> see the value).
>>>>>
>>>>> GitHub link:
>>>>>
>>>>
>>> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>>>>>
>>>>> ----
>>>>> This is an automatically sent email for dev@ossie.apache.org.
>>>>> To unsubscribe, please send an email to:
>>>> dev-unsubscribe@ossie.apache.org
>>>>>
>>>>>
>>>>
>>>
>


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Kyoung Min Kim <ka...@gmail.com>.
A data point on Chris's question about the input shape, from an experiment I
ran this week on real tables.

I took one database of 33 datasets, most of them monthly profile marts, with
45 relationships, generated an Ossie model for it, and gave three models
(one current, two older or smaller) a catalog of the datasets, fields and
three named metrics. Their only job was to return a spec in the Layer 3
shape: a metric or an aggregate, dimensions, filters. A compiler resolved
the joins from the relationships and produced the SQL. I scored ten
questions by compiling the reference spec and the model's spec and comparing
the SQL, counting result-equivalent variants as correct, and then ran the
pairs that differed against the warehouse to check those equivalence calls.
The current model got 8 of 10, the other two 7 and 8.

What made the difference was not the models but what the compiler could
check because the input was tied to the Ossie model. When a model picked a
dataset with no relationship path to the others, compilation refused instead
of producing a cross join. A filter on the period column of one monthly
dataset could be carried to the joined one, so the one spec that dropped it
did not count across months. And where a dataset declared unique_keys,
COUNT(*) against COUNT(DISTINCT key) could be proven equivalent or not,
which the warehouse run confirmed in both directions. The keys also
explained a mistake of my own: an alternative answer I had accepted counted
distinct values of a column that is only part of the declared key, and on
real data it undercounted by more than ten times. All of that is much harder
to offer once the server receives an engine-native string.

So I would keep a Layer 3 style input as the portable path and treat the
native-query input in #529 as the escape hatch Chris describes, with the
model identity required rather than optional whenever the request claims to
be about an Ossie model.

One more thing that touches Markus's filter value alternatives. The most
common remaining mistake involved code values: inventing a label for a
job-segment column, or not knowing that a filter on a code column was needed
at all, because nothing in the model lists the admissible values. The
ontology spec already has a place for that in a ValueType's requires list.
If execute_query returns value alternatives after the fact, it is worth
deciding whether that vocabulary also belongs in the model, where the agent
sees it before it queries.

Kyoung Min Kim


> 2026. 10. 7. AM 8:51, Aleksandr Kosachev <a....@gmail.com> wrote:
> 
> Hi Chris,
> 
> Two things would make it easier:
> 
> - The spec should require a core entry in the document, not only support for it in the reader.
> - Whatever changes a query result should not sit in custom_extensions, which the spec calls metadata and which a reader from another vendor cannot interpret.
> 
> === Why the first ===
> 
> I ran the question against what the converters produce on their own vendor-format inputs, plus the examples: current main, 19 documents, 411 field and metric expressions. 32 of them carry a vendor dialect only.
> 
> The 21 tagged DATABRICKS are plain SQL: bare columns, SUM, COUNT(*), COUNT(DISTINCT ...). ossie-clickhouse parses them unchanged and still cannot use them, because nothing in the document says this text is also core, and the spec lets a reader ignore other dialects. Every expression in those documents is affected, so hiding them would leave nothing to serve. I could transpile DATABRICKS with SQLGlot, and for these 21 it would work. But then a reader relies on the vendor entry and the core entry becomes the optional one, the opposite of what the spec says.
> 
> The other 11 are THOUGHTSPOT, and they show where the gap is. In the TPC-DS fixture every field carries ANSI_SQL next to its formula; none of the 10 metrics does. The fixture is small, but #462 measured the same thing on 31 real ThoughtSpot models: 476 of 476 metrics have no portable dialect, and the Snowflake and Databricks converters drop them on the way out.
> 
> Reading the formula as SQL anyway is worse than refusing it. 9 of the 11 go through a SQL parser without an error, with [store_sales::ss_ext_sales_price] coming out as a CAST, so the mistake would show up in the numbers. "Other dialects MAY be ignored" is the right rule as long as there is a core entry to fall back on.
> 
> What it costs:
> 
> Databricks: for these 21, one more line with the same text.
> ThoughtSpot: the measure column declared as a field, and the metric written over it. #463 tried SUM(dataset.column) and withdrew it, since readers split on whether that names a field or a column. #486 settles that it names a field.
> If a required core entry is too much for fields, metrics alone would close the gap. A weaker form would be a machine-readable statement that an expression has no core form, so a compiler can hide one metric instead of refusing the document.
> 
> === Why the second ===
> 
> Of the 79 custom_extensions entries, most are display formats, connection names and round-trip witnesses. One kind is different: the Databricks fixtures keep a model-level filter there (ss_net_profit > 0), and the core has no place for it. A reader from another vendor ignores the extension, so on the same document it returns different numbers from Databricks, and nothing in the document warns about that.
> 
> The multiplicity thread is moving relationship cardinality into core. A model filter deserves the same question. For the rest, the per-converter schema documentation you +1'd would at least let a compiler see what it is ignoring.
> 
> Kind regards, 
> Alex
> 
>> On 6 Oct 2026, at 14.52, Chris Eubank via dev <de...@ossie.apache.org> wrote:
>> 
>> Hi Alex,
>> 
>> I think we'll want to support both query interfaces wherever possible. #452
>> brings up different cases where one layer is more natural. For agents, I'd
>> expect both would be useful: layer 3 for handling many common analyses, and
>> layer 2 for when the agent wants to perform analysis not directly supported
>> by the model (e.g. joining in another table, tweaking a relationship) while
>> still having some guardrails (e.g. keeping the measure abstraction vs.
>> falling back to raw SQL).
>> 
>> It is worth noting a few limitations: Ossie currently lacks official query
>> semantics. #452 points to the layered PRs, but those are still in flight.
>> There are also some foundational data modeling PRs (#343, #486) that would
>> be great to land, since some core behavior (e.g. name resolution rules)
>> remains underspecified.
>> 
>> I would be really interested in your perspective on how you think about
>> querying Ossie models, especially since you're building on top of the
>> project. In this thread
>> <https://lists.apache.org/thread/r1vhmw2vdms3bvv1s9tsfm06knnnbn30> we were
>> discussing the possibility of vendor-specific extensions (both in
>> custom_extensions and expressions) being made more transparent. The
>> expression spec only states that compliant implementations must support
>> Ossie SQL, but I suspect in practice many of these vendor extensions will
>> exist in models. What would make your job easier when building your query
>> compilation on top of real-world models?
>> 
>> Kind regards,
>> Chris
>> 
>> 
>> On Mon, Oct 5, 2026 at 4:23 PM Aleksandr Kosachev <a....@gmail.com>
>> wrote:
>> 
>>> Hi Markus, JB,
>>> 
>>> One gap the thread has not named yet. The REST API spec proposed on 09-23
>>> covers two things: managing models and binding a model to physical
>>> entities. It does not cover running a query and returning data. That is
>>> where execute_query starts. So an MCP spec next to the REST spec would not
>>> duplicate it. It would fill the missing step.
>>> 
>>> Markus asked how an agent that found a model knows how to run a query
>>> against the source.
>>> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
>>> does not need to know:
>>> - The model carries no connection.
>>> - The server starts with a model and a database connection.
>>> - The agent sees only the tools.
>>> 
>>> This reconciles the two positions in the thread. Ossie specifies the tool
>>> contract. The engine or catalog that serves it owns the binding.
>>> 
>>> One question needs an answer before a spec: what is the input of
>>> execute_query? PR #452 names two query interfaces:
>>> - Layer 2: SQL with MEASURE().
>>> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>>> 
>>> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
>>> When its planner refuses a query, the refusal names the cause.
>>> For example, "unknown metric 'x' (did you mean: ...)", or which metric
>>> would be counted once per joined row.
>>> The agent reads that and repairs the query.
>>> Markus listed four outputs: data, errors, suggestions, filter value
>>> alternatives.
>>> ossie-clickhouse returns data, errors and name suggestions, plus the
>>> generated SQL. It has no filter value alternatives yet.
>>> I would like to hear how those look on your side.
>>> 
>>> On query_model and the harness dependency, ossie-clickhouse confirms it.
>>> Its overview tool returns datasets and metrics without fields.
>>> Agents then call a search tool and a describe-one-object tool before every
>>> query. A harness with the file on disk, as Markus described, does not need
>>> them. So discovery tools belong in the spec as optional, not dropped.
>>> 
>>> Warm regards,
>>> Alex Kosachev
>>> 
>>> [1] https://github.com/Dactopus/ossie-clickhouse
>>> <
>>> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
>>>> 
>>> 
>>> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>>> 
>>>> I'm curious about the expectations here. If e.g. GitHub copilot discovers
>>>> an Ossie semantic model, how would it know how to connect/execute a query
>>>> against the data source (e.g. PBI semantic model, Databricks, Snowflake,
>>>> ...)? Is the expectation that the harness would get additional
>>> instructions
>>>> anyway?
>>>> 
>>>> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
>>>> implementation.
>>>> ________________________________
>>>> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
>>>> Sent: Monday, October 5, 2026 8:25 PM
>>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>>> live) [ossie]
>>>> 
>>>> Hi
>>>> 
>>>> It's not clear to me if we are discussing the specification or the
>>>> implementation. MCP implementation in Ossie: I'm not convinced; it seems
>>>> more like an engine/catalog implementation.
>>>> The spec could make sense.
>>>> 
>>>> But do we really need the MCP spec in Ossie for agent model discovery?
>>> I'm
>>>> not sure. That's actually one of the reasons for the REST API Spec too
>>>> (model discovery, versioning, etc).
>>>> We could have the MCP spec alongside the REST API Spec.
>>>> 
>>>> Also, MCP is good to have but clearly not a requirement for the agent
>>> (API
>>>> or CLI are classic data providers for agent, including operations).
>>>> 
>>>> Regards
>>>> JB
>>>> 
>>>> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>>>> 
>>>>> I'd like to provide a slightly different pitch for an MCP server spec
>>>> (not
>>>>> necessarily the implementation). From my perspective, Ossies biggest
>>>> value
>>>>> prop is related to empowering agents to understand semantics and
>>>>> subsequently build queries to retrieve data.
>>>>> Most important is probably an execute_query MCP tool.
>>>>> 
>>>>> 
>>>>> *
>>>>> Input: data source id, query
>>>>> *
>>>>> Output
>>>>>    *
>>>>> Data
>>>>>    *
>>>>> Errors
>>>>>    *
>>>>> Suggestions
>>>>>    *
>>>>> Filter value alternatives
>>>>>    *
>>>>> ...
>>>>> 
>>>>> This operates under the assumption that the agent is able to
>>> discover/see
>>>>> the Ossie semantic model (e.g. using
>>>> http://agenticresourcediscovery.org/).
>>>>> We already have a few variants of execute_query tool across our
>>> engines.
>>>>> 
>>>>> I'm not sure about query_model (yet), at least for now we see a
>>>> dependency
>>>>> on the agent harness. Some harness can only digest a small view of the
>>>>> semantic model and thus need tools to progressively discover, others
>>>> (like
>>>>> e.g. GitHub copilot cli), don't need any of that, as they'll have full
>>>>> filesystem access, download/query local.
>>>>> 
>>>>> ________________________________
>>>>> From: jbonofre (via GitHub) <gi...@apache.org>
>>>>> Sent: Monday, October 5, 2026 8:01 AM
>>>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>>>> live) [ossie]
>>>>> 
>>>>> 
>>>>> GitHub user jbonofre added a comment to the discussion: Proposal: an
>>> MCP
>>>>> server for Ossie (and where it should live)
>>>>> 
>>>>> I'm not sure we need MCP server at all in Ossie. That could be
>>> something
>>>>> catalogs can do.
>>>>> 
>>>>> I would rather not consider MCP service directly in Ossie for now (I
>>>> don't
>>>>> see the value).
>>>>> 
>>>>> GitHub link:
>>>>> 
>>>> 
>>> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>>>>> 
>>>>> ----
>>>>> This is an automatically sent email for dev@ossie.apache.org.
>>>>> To unsubscribe, please send an email to:
>>>> dev-unsubscribe@ossie.apache.org
>>>>> 
>>>>> 
>>>> 
>>> 
> 


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Chris Eubank via dev <de...@ossie.apache.org>.
Hi Alex,

Thanks for the analysis. So far, most of the repo has focused on enabling
documents to move between vendors; now that we are pushing portable query
semantics and executing against real models, these are exactly the gaps we
should close.

On the first point, I agree: documents should contain a required portable
core representation. I would not rely on readers transpiling vendor
dialects as the general solution. Not every vendor expression language is
SQL, and, as you point out, treating a non-SQL formula as SQL can parse
successfully while changing the result, which is worse than refusing it. To
your point, I suspect there will be a period where Ossie query implementers
drop measures / fields not expressible in the portable SQL.

For Databricks, I agree we should emit a portable SQL expression wherever
one exists. ThoughtSpot is a useful example of the harder case: some of its
measures reduce to a simple aggregate over a single column and translate
cleanly (they export both their dialect & SQL), while others use more
complex formula constructs. Some exporters carry all measures only in their
vendor dialect, with no portable form at all.

On the second point, I also agree that anything which changes query results
should not live only in custom_extensions.

For the Databricks metric view filters specifically, though, I expect the
converter to change rather than the spec to add a new core field for it.
With the layering work, a metric view looks more like a dataset or table
than a full semantic model. A filter that defines the dataset can therefore
be represented in the source relation itself, rather than carried as opaque
metadata. That puts the semantics in the queryable core and removes the
divergence you identified. Some vendors support model-level filters so this
could very well be something added to the core in the future.

Lest it get lost since this thread forked, I'm curious about your thoughts
on this part
<https://lists.apache.org/thread/y7zdqybc1c83hfovqmlp15fnh756m0do>of the
thread regarding MCP operability.

Kind regards,
Chris

On Tue, Oct 6, 2026 at 4:51 PM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi Chris,
>
> Two things would make it easier:
>
> - The spec should require a core entry in the document, not only support
> for it in the reader.
> - Whatever changes a query result should not sit in custom_extensions,
> which the spec calls metadata and which a reader from another vendor cannot
> interpret.
>
> === Why the first ===
>
> I ran the question against what the converters produce on their own
> vendor-format inputs, plus the examples: current main, 19 documents, 411
> field and metric expressions. 32 of them carry a vendor dialect only.
>
> The 21 tagged DATABRICKS are plain SQL: bare columns, SUM, COUNT(*),
> COUNT(DISTINCT ...). ossie-clickhouse parses them unchanged and still
> cannot use them, because nothing in the document says this text is also
> core, and the spec lets a reader ignore other dialects. Every expression in
> those documents is affected, so hiding them would leave nothing to serve. I
> could transpile DATABRICKS with SQLGlot, and for these 21 it would work.
> But then a reader relies on the vendor entry and the core entry becomes the
> optional one, the opposite of what the spec says.
>
> The other 11 are THOUGHTSPOT, and they show where the gap is. In the
> TPC-DS fixture every field carries ANSI_SQL next to its formula; none of
> the 10 metrics does. The fixture is small, but #462 measured the same thing
> on 31 real ThoughtSpot models: 476 of 476 metrics have no portable dialect,
> and the Snowflake and Databricks converters drop them on the way out.
>
> Reading the formula as SQL anyway is worse than refusing it. 9 of the 11
> go through a SQL parser without an error, with
> [store_sales::ss_ext_sales_price] coming out as a CAST, so the mistake
> would show up in the numbers. "Other dialects MAY be ignored" is the right
> rule as long as there is a core entry to fall back on.
>
> What it costs:
>
> Databricks: for these 21, one more line with the same text.
> ThoughtSpot: the measure column declared as a field, and the metric
> written over it. #463 tried SUM(dataset.column) and withdrew it, since
> readers split on whether that names a field or a column. #486 settles that
> it names a field.
> If a required core entry is too much for fields, metrics alone would close
> the gap. A weaker form would be a machine-readable statement that an
> expression has no core form, so a compiler can hide one metric instead of
> refusing the document.
>
> === Why the second ===
>
> Of the 79 custom_extensions entries, most are display formats, connection
> names and round-trip witnesses. One kind is different: the Databricks
> fixtures keep a model-level filter there (ss_net_profit > 0), and the core
> has no place for it. A reader from another vendor ignores the extension, so
> on the same document it returns different numbers from Databricks, and
> nothing in the document warns about that.
>
> The multiplicity thread is moving relationship cardinality into core. A
> model filter deserves the same question. For the rest, the per-converter
> schema documentation you +1'd would at least let a compiler see what it is
> ignoring.
>
> Kind regards,
> Alex
>
> > On 6 Oct 2026, at 14.52, Chris Eubank via dev <de...@ossie.apache.org>
> wrote:
> >
> > Hi Alex,
> >
> > I think we'll want to support both query interfaces wherever possible.
> #452
> > brings up different cases where one layer is more natural. For agents,
> I'd
> > expect both would be useful: layer 3 for handling many common analyses,
> and
> > layer 2 for when the agent wants to perform analysis not directly
> supported
> > by the model (e.g. joining in another table, tweaking a relationship)
> while
> > still having some guardrails (e.g. keeping the measure abstraction vs.
> > falling back to raw SQL).
> >
> > It is worth noting a few limitations: Ossie currently lacks official
> query
> > semantics. #452 points to the layered PRs, but those are still in flight.
> > There are also some foundational data modeling PRs (#343, #486) that
> would
> > be great to land, since some core behavior (e.g. name resolution rules)
> > remains underspecified.
> >
> > I would be really interested in your perspective on how you think about
> > querying Ossie models, especially since you're building on top of the
> > project. In this thread
> > <https://lists.apache.org/thread/r1vhmw2vdms3bvv1s9tsfm06knnnbn30> we
> were
> > discussing the possibility of vendor-specific extensions (both in
> > custom_extensions and expressions) being made more transparent. The
> > expression spec only states that compliant implementations must support
> > Ossie SQL, but I suspect in practice many of these vendor extensions will
> > exist in models. What would make your job easier when building your query
> > compilation on top of real-world models?
> >
> > Kind regards,
> > Chris
> >
> >
> > On Mon, Oct 5, 2026 at 4:23 PM Aleksandr Kosachev <a.yakutskiy@gmail.com
> >
> > wrote:
> >
> >> Hi Markus, JB,
> >>
> >> One gap the thread has not named yet. The REST API spec proposed on
> 09-23
> >> covers two things: managing models and binding a model to physical
> >> entities. It does not cover running a query and returning data. That is
> >> where execute_query starts. So an MCP spec next to the REST spec would
> not
> >> duplicate it. It would fill the missing step.
> >>
> >> Markus asked how an agent that found a model knows how to run a query
> >> against the source.
> >> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain,
> it
> >> does not need to know:
> >> - The model carries no connection.
> >> - The server starts with a model and a database connection.
> >> - The agent sees only the tools.
> >>
> >> This reconciles the two positions in the thread. Ossie specifies the
> tool
> >> contract. The engine or catalog that serves it owns the binding.
> >>
> >> One question needs an answer before a spec: what is the input of
> >> execute_query? PR #452 names two query interfaces:
> >> - Layer 2: SQL with MEASURE().
> >> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
> >>
> >> execute_query could accept either. ossie-clickhouse takes the Layer 3
> form.
> >> When its planner refuses a query, the refusal names the cause.
> >> For example, "unknown metric 'x' (did you mean: ...)", or which metric
> >> would be counted once per joined row.
> >> The agent reads that and repairs the query.
> >> Markus listed four outputs: data, errors, suggestions, filter value
> >> alternatives.
> >> ossie-clickhouse returns data, errors and name suggestions, plus the
> >> generated SQL. It has no filter value alternatives yet.
> >> I would like to hear how those look on your side.
> >>
> >> On query_model and the harness dependency, ossie-clickhouse confirms it.
> >> Its overview tool returns datasets and metrics without fields.
> >> Agents then call a search tool and a describe-one-object tool before
> every
> >> query. A harness with the file on disk, as Markus described, does not
> need
> >> them. So discovery tools belong in the spec as optional, not dropped.
> >>
> >> Warm regards,
> >> Alex Kosachev
> >>
> >> [1] https://github.com/Dactopus/ossie-clickhouse
> >> <
> >>
> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
> >>>
> >>
> >> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
> >>
> >>> I'm curious about the expectations here. If e.g. GitHub copilot
> discovers
> >>> an Ossie semantic model, how would it know how to connect/execute a
> query
> >>> against the data source (e.g. PBI semantic model, Databricks,
> Snowflake,
> >>> ...)? Is the expectation that the harness would get additional
> >> instructions
> >>> anyway?
> >>>
> >>> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> >>> implementation.
> >>> ________________________________
> >>> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> >>> Sent: Monday, October 5, 2026 8:25 PM
> >>> To: dev@ossie.apache.org <de...@ossie.apache.org>
> >>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> >>> live) [ossie]
> >>>
> >>> Hi
> >>>
> >>> It's not clear to me if we are discussing the specification or the
> >>> implementation. MCP implementation in Ossie: I'm not convinced; it
> seems
> >>> more like an engine/catalog implementation.
> >>> The spec could make sense.
> >>>
> >>> But do we really need the MCP spec in Ossie for agent model discovery?
> >> I'm
> >>> not sure. That's actually one of the reasons for the REST API Spec too
> >>> (model discovery, versioning, etc).
> >>> We could have the MCP spec alongside the REST API Spec.
> >>>
> >>> Also, MCP is good to have but clearly not a requirement for the agent
> >> (API
> >>> or CLI are classic data providers for agent, including operations).
> >>>
> >>> Regards
> >>> JB
> >>>
> >>> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> >>>
> >>>> I'd like to provide a slightly different pitch for an MCP server spec
> >>> (not
> >>>> necessarily the implementation). From my perspective, Ossies biggest
> >>> value
> >>>> prop is related to empowering agents to understand semantics and
> >>>> subsequently build queries to retrieve data.
> >>>> Most important is probably an execute_query MCP tool.
> >>>>
> >>>>
> >>>>  *
> >>>> Input: data source id, query
> >>>>  *
> >>>> Output
> >>>>     *
> >>>> Data
> >>>>     *
> >>>> Errors
> >>>>     *
> >>>> Suggestions
> >>>>     *
> >>>> Filter value alternatives
> >>>>     *
> >>>> ...
> >>>>
> >>>> This operates under the assumption that the agent is able to
> >> discover/see
> >>>> the Ossie semantic model (e.g. using
> >>> http://agenticresourcediscovery.org/).
> >>>> We already have a few variants of execute_query tool across our
> >> engines.
> >>>>
> >>>> I'm not sure about query_model (yet), at least for now we see a
> >>> dependency
> >>>> on the agent harness. Some harness can only digest a small view of the
> >>>> semantic model and thus need tools to progressively discover, others
> >>> (like
> >>>> e.g. GitHub copilot cli), don't need any of that, as they'll have full
> >>>> filesystem access, download/query local.
> >>>>
> >>>> ________________________________
> >>>> From: jbonofre (via GitHub) <gi...@apache.org>
> >>>> Sent: Monday, October 5, 2026 8:01 AM
> >>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
> >>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
> should
> >>>> live) [ossie]
> >>>>
> >>>>
> >>>> GitHub user jbonofre added a comment to the discussion: Proposal: an
> >> MCP
> >>>> server for Ossie (and where it should live)
> >>>>
> >>>> I'm not sure we need MCP server at all in Ossie. That could be
> >> something
> >>>> catalogs can do.
> >>>>
> >>>> I would rather not consider MCP service directly in Ossie for now (I
> >>> don't
> >>>> see the value).
> >>>>
> >>>> GitHub link:
> >>>>
> >>>
> >>
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> >>>>
> >>>> ----
> >>>> This is an automatically sent email for dev@ossie.apache.org.
> >>>> To unsubscribe, please send an email to:
> >>> dev-unsubscribe@ossie.apache.org
> >>>>
> >>>>
> >>>
> >>
>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi Chris,

Two things would make it easier:

- The spec should require a core entry in the document, not only support for it in the reader.
- Whatever changes a query result should not sit in custom_extensions, which the spec calls metadata and which a reader from another vendor cannot interpret.

=== Why the first ===

I ran the question against what the converters produce on their own vendor-format inputs, plus the examples: current main, 19 documents, 411 field and metric expressions. 32 of them carry a vendor dialect only.

The 21 tagged DATABRICKS are plain SQL: bare columns, SUM, COUNT(*), COUNT(DISTINCT ...). ossie-clickhouse parses them unchanged and still cannot use them, because nothing in the document says this text is also core, and the spec lets a reader ignore other dialects. Every expression in those documents is affected, so hiding them would leave nothing to serve. I could transpile DATABRICKS with SQLGlot, and for these 21 it would work. But then a reader relies on the vendor entry and the core entry becomes the optional one, the opposite of what the spec says.

The other 11 are THOUGHTSPOT, and they show where the gap is. In the TPC-DS fixture every field carries ANSI_SQL next to its formula; none of the 10 metrics does. The fixture is small, but #462 measured the same thing on 31 real ThoughtSpot models: 476 of 476 metrics have no portable dialect, and the Snowflake and Databricks converters drop them on the way out.

Reading the formula as SQL anyway is worse than refusing it. 9 of the 11 go through a SQL parser without an error, with [store_sales::ss_ext_sales_price] coming out as a CAST, so the mistake would show up in the numbers. "Other dialects MAY be ignored" is the right rule as long as there is a core entry to fall back on.

What it costs:

Databricks: for these 21, one more line with the same text.
ThoughtSpot: the measure column declared as a field, and the metric written over it. #463 tried SUM(dataset.column) and withdrew it, since readers split on whether that names a field or a column. #486 settles that it names a field.
If a required core entry is too much for fields, metrics alone would close the gap. A weaker form would be a machine-readable statement that an expression has no core form, so a compiler can hide one metric instead of refusing the document.

=== Why the second ===

Of the 79 custom_extensions entries, most are display formats, connection names and round-trip witnesses. One kind is different: the Databricks fixtures keep a model-level filter there (ss_net_profit > 0), and the core has no place for it. A reader from another vendor ignores the extension, so on the same document it returns different numbers from Databricks, and nothing in the document warns about that.

The multiplicity thread is moving relationship cardinality into core. A model filter deserves the same question. For the rest, the per-converter schema documentation you +1'd would at least let a compiler see what it is ignoring.

Kind regards, 
Alex

> On 6 Oct 2026, at 14.52, Chris Eubank via dev <de...@ossie.apache.org> wrote:
> 
> Hi Alex,
> 
> I think we'll want to support both query interfaces wherever possible. #452
> brings up different cases where one layer is more natural. For agents, I'd
> expect both would be useful: layer 3 for handling many common analyses, and
> layer 2 for when the agent wants to perform analysis not directly supported
> by the model (e.g. joining in another table, tweaking a relationship) while
> still having some guardrails (e.g. keeping the measure abstraction vs.
> falling back to raw SQL).
> 
> It is worth noting a few limitations: Ossie currently lacks official query
> semantics. #452 points to the layered PRs, but those are still in flight.
> There are also some foundational data modeling PRs (#343, #486) that would
> be great to land, since some core behavior (e.g. name resolution rules)
> remains underspecified.
> 
> I would be really interested in your perspective on how you think about
> querying Ossie models, especially since you're building on top of the
> project. In this thread
> <https://lists.apache.org/thread/r1vhmw2vdms3bvv1s9tsfm06knnnbn30> we were
> discussing the possibility of vendor-specific extensions (both in
> custom_extensions and expressions) being made more transparent. The
> expression spec only states that compliant implementations must support
> Ossie SQL, but I suspect in practice many of these vendor extensions will
> exist in models. What would make your job easier when building your query
> compilation on top of real-world models?
> 
> Kind regards,
> Chris
> 
> 
> On Mon, Oct 5, 2026 at 4:23 PM Aleksandr Kosachev <a....@gmail.com>
> wrote:
> 
>> Hi Markus, JB,
>> 
>> One gap the thread has not named yet. The REST API spec proposed on 09-23
>> covers two things: managing models and binding a model to physical
>> entities. It does not cover running a query and returning data. That is
>> where execute_query starts. So an MCP spec next to the REST spec would not
>> duplicate it. It would fill the missing step.
>> 
>> Markus asked how an agent that found a model knows how to run a query
>> against the source.
>> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
>> does not need to know:
>> - The model carries no connection.
>> - The server starts with a model and a database connection.
>> - The agent sees only the tools.
>> 
>> This reconciles the two positions in the thread. Ossie specifies the tool
>> contract. The engine or catalog that serves it owns the binding.
>> 
>> One question needs an answer before a spec: what is the input of
>> execute_query? PR #452 names two query interfaces:
>> - Layer 2: SQL with MEASURE().
>> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>> 
>> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
>> When its planner refuses a query, the refusal names the cause.
>> For example, "unknown metric 'x' (did you mean: ...)", or which metric
>> would be counted once per joined row.
>> The agent reads that and repairs the query.
>> Markus listed four outputs: data, errors, suggestions, filter value
>> alternatives.
>> ossie-clickhouse returns data, errors and name suggestions, plus the
>> generated SQL. It has no filter value alternatives yet.
>> I would like to hear how those look on your side.
>> 
>> On query_model and the harness dependency, ossie-clickhouse confirms it.
>> Its overview tool returns datasets and metrics without fields.
>> Agents then call a search tool and a describe-one-object tool before every
>> query. A harness with the file on disk, as Markus described, does not need
>> them. So discovery tools belong in the spec as optional, not dropped.
>> 
>> Warm regards,
>> Alex Kosachev
>> 
>> [1] https://github.com/Dactopus/ossie-clickhouse
>> <
>> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
>>> 
>> 
>> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>> 
>>> I'm curious about the expectations here. If e.g. GitHub copilot discovers
>>> an Ossie semantic model, how would it know how to connect/execute a query
>>> against the data source (e.g. PBI semantic model, Databricks, Snowflake,
>>> ...)? Is the expectation that the harness would get additional
>> instructions
>>> anyway?
>>> 
>>> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
>>> implementation.
>>> ________________________________
>>> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
>>> Sent: Monday, October 5, 2026 8:25 PM
>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>> live) [ossie]
>>> 
>>> Hi
>>> 
>>> It's not clear to me if we are discussing the specification or the
>>> implementation. MCP implementation in Ossie: I'm not convinced; it seems
>>> more like an engine/catalog implementation.
>>> The spec could make sense.
>>> 
>>> But do we really need the MCP spec in Ossie for agent model discovery?
>> I'm
>>> not sure. That's actually one of the reasons for the REST API Spec too
>>> (model discovery, versioning, etc).
>>> We could have the MCP spec alongside the REST API Spec.
>>> 
>>> Also, MCP is good to have but clearly not a requirement for the agent
>> (API
>>> or CLI are classic data providers for agent, including operations).
>>> 
>>> Regards
>>> JB
>>> 
>>> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>>> 
>>>> I'd like to provide a slightly different pitch for an MCP server spec
>>> (not
>>>> necessarily the implementation). From my perspective, Ossies biggest
>>> value
>>>> prop is related to empowering agents to understand semantics and
>>>> subsequently build queries to retrieve data.
>>>> Most important is probably an execute_query MCP tool.
>>>> 
>>>> 
>>>>  *
>>>> Input: data source id, query
>>>>  *
>>>> Output
>>>>     *
>>>> Data
>>>>     *
>>>> Errors
>>>>     *
>>>> Suggestions
>>>>     *
>>>> Filter value alternatives
>>>>     *
>>>> ...
>>>> 
>>>> This operates under the assumption that the agent is able to
>> discover/see
>>>> the Ossie semantic model (e.g. using
>>> http://agenticresourcediscovery.org/).
>>>> We already have a few variants of execute_query tool across our
>> engines.
>>>> 
>>>> I'm not sure about query_model (yet), at least for now we see a
>>> dependency
>>>> on the agent harness. Some harness can only digest a small view of the
>>>> semantic model and thus need tools to progressively discover, others
>>> (like
>>>> e.g. GitHub copilot cli), don't need any of that, as they'll have full
>>>> filesystem access, download/query local.
>>>> 
>>>> ________________________________
>>>> From: jbonofre (via GitHub) <gi...@apache.org>
>>>> Sent: Monday, October 5, 2026 8:01 AM
>>>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>>>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>>>> live) [ossie]
>>>> 
>>>> 
>>>> GitHub user jbonofre added a comment to the discussion: Proposal: an
>> MCP
>>>> server for Ossie (and where it should live)
>>>> 
>>>> I'm not sure we need MCP server at all in Ossie. That could be
>> something
>>>> catalogs can do.
>>>> 
>>>> I would rather not consider MCP service directly in Ossie for now (I
>>> don't
>>>> see the value).
>>>> 
>>>> GitHub link:
>>>> 
>>> 
>> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>>>> 
>>>> ----
>>>> This is an automatically sent email for dev@ossie.apache.org.
>>>> To unsubscribe, please send an email to:
>>> dev-unsubscribe@ossie.apache.org
>>>> 
>>>> 
>>> 
>> 


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Chris Eubank via dev <de...@ossie.apache.org>.
Hi Alex,

I think we'll want to support both query interfaces wherever possible. #452
brings up different cases where one layer is more natural. For agents, I'd
expect both would be useful: layer 3 for handling many common analyses, and
layer 2 for when the agent wants to perform analysis not directly supported
by the model (e.g. joining in another table, tweaking a relationship) while
still having some guardrails (e.g. keeping the measure abstraction vs.
falling back to raw SQL).

It is worth noting a few limitations: Ossie currently lacks official query
semantics. #452 points to the layered PRs, but those are still in flight.
There are also some foundational data modeling PRs (#343, #486) that would
be great to land, since some core behavior (e.g. name resolution rules)
remains underspecified.

I would be really interested in your perspective on how you think about
querying Ossie models, especially since you're building on top of the
project. In this thread
<https://lists.apache.org/thread/r1vhmw2vdms3bvv1s9tsfm06knnnbn30> we were
discussing the possibility of vendor-specific extensions (both in
custom_extensions and expressions) being made more transparent. The
expression spec only states that compliant implementations must support
Ossie SQL, but I suspect in practice many of these vendor extensions will
exist in models. What would make your job easier when building your query
compilation on top of real-world models?

Kind regards,
Chris


On Mon, Oct 5, 2026 at 4:23 PM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi Markus, JB,
>
> One gap the thread has not named yet. The REST API spec proposed on 09-23
> covers two things: managing models and binding a model to physical
> entities. It does not cover running a query and returning data. That is
> where execute_query starts. So an MCP spec next to the REST spec would not
> duplicate it. It would fill the missing step.
>
> Markus asked how an agent that found a model knows how to run a query
> against the source.
> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
> does not need to know:
> - The model carries no connection.
> - The server starts with a model and a database connection.
> - The agent sees only the tools.
>
> This reconciles the two positions in the thread. Ossie specifies the tool
> contract. The engine or catalog that serves it owns the binding.
>
> One question needs an answer before a spec: what is the input of
> execute_query? PR #452 names two query interfaces:
> - Layer 2: SQL with MEASURE().
> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>
> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
> When its planner refuses a query, the refusal names the cause.
> For example, "unknown metric 'x' (did you mean: ...)", or which metric
> would be counted once per joined row.
> The agent reads that and repairs the query.
> Markus listed four outputs: data, errors, suggestions, filter value
> alternatives.
> ossie-clickhouse returns data, errors and name suggestions, plus the
> generated SQL. It has no filter value alternatives yet.
> I would like to hear how those look on your side.
>
> On query_model and the harness dependency, ossie-clickhouse confirms it.
> Its overview tool returns datasets and metrics without fields.
> Agents then call a search tool and a describe-one-object tool before every
> query. A harness with the file on disk, as Markus described, does not need
> them. So discovery tools belong in the spec as optional, not dropped.
>
> Warm regards,
> Alex Kosachev
>
> [1] https://github.com/Dactopus/ossie-clickhouse
> <
> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
> >
>
> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > I'm curious about the expectations here. If e.g. GitHub copilot discovers
> > an Ossie semantic model, how would it know how to connect/execute a query
> > against the data source (e.g. PBI semantic model, Databricks, Snowflake,
> > ...)? Is the expectation that the harness would get additional
> instructions
> > anyway?
> >
> > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> > implementation.
> > ________________________________
> > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> > Sent: Monday, October 5, 2026 8:25 PM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> > Hi
> >
> > It's not clear to me if we are discussing the specification or the
> > implementation. MCP implementation in Ossie: I'm not convinced; it seems
> > more like an engine/catalog implementation.
> > The spec could make sense.
> >
> > But do we really need the MCP spec in Ossie for agent model discovery?
> I'm
> > not sure. That's actually one of the reasons for the REST API Spec too
> > (model discovery, versioning, etc).
> > We could have the MCP spec alongside the REST API Spec.
> >
> > Also, MCP is good to have but clearly not a requirement for the agent
> (API
> > or CLI are classic data providers for agent, including operations).
> >
> > Regards
> > JB
> >
> > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> >
> > > I'd like to provide a slightly different pitch for an MCP server spec
> > (not
> > > necessarily the implementation). From my perspective, Ossies biggest
> > value
> > > prop is related to empowering agents to understand semantics and
> > > subsequently build queries to retrieve data.
> > > Most important is probably an execute_query MCP tool.
> > >
> > >
> > >   *
> > > Input: data source id, query
> > >   *
> > > Output
> > >      *
> > > Data
> > >      *
> > > Errors
> > >      *
> > > Suggestions
> > >      *
> > > Filter value alternatives
> > >      *
> > > ...
> > >
> > > This operates under the assumption that the agent is able to
> discover/see
> > > the Ossie semantic model (e.g. using
> > http://agenticresourcediscovery.org/).
> > > We already have a few variants of execute_query tool across our
> engines.
> > >
> > > I'm not sure about query_model (yet), at least for now we see a
> > dependency
> > > on the agent harness. Some harness can only digest a small view of the
> > > semantic model and thus need tools to progressively discover, others
> > (like
> > > e.g. GitHub copilot cli), don't need any of that, as they'll have full
> > > filesystem access, download/query local.
> > >
> > > ________________________________
> > > From: jbonofre (via GitHub) <gi...@apache.org>
> > > Sent: Monday, October 5, 2026 8:01 AM
> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > > live) [ossie]
> > >
> > >
> > > GitHub user jbonofre added a comment to the discussion: Proposal: an
> MCP
> > > server for Ossie (and where it should live)
> > >
> > > I'm not sure we need MCP server at all in Ossie. That could be
> something
> > > catalogs can do.
> > >
> > > I would rather not consider MCP service directly in Ossie for now (I
> > don't
> > > see the value).
> > >
> > > GitHub link:
> > >
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> > >
> > > ----
> > > This is an automatically sent email for dev@ossie.apache.org.
> > > To unsubscribe, please send an email to:
> > dev-unsubscribe@ossie.apache.org
> > >
> > >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Chris Eubank via dev <de...@ossie.apache.org>.
JB,

Agreed: the dividing line is whether the query resolves against an Ossie
model.

From the Databricks side, there's interest in building Ossie query ->
SQL/data capabilities (REST, functions, MCP, etc) on top of the Layer 2 and
Layer 3 query interfaces, where the query resolves against the model. The
prerequisite is landing those interfaces in the core spec. Ossie models are
not queryable under the spec today: the query interface is still a roadmap
item, and the expression-language spec itself says that "The SQL Interface
will be defined in a different document."

Markus, JB,

I also think we should align on what vendor-specific dialects mean. Today
the spec requires support for models using the portable Ossie subset and
treats other dialects as optional expression-level extensions: "The Ossie
dialect should always be supported. Other dialects MAY be ignored. There is
no guarantee that all different dialects for an expression will act the
same."

I've understood those dialects as a round-tripping mechanism for vendor <->
Ossie fidelity, not as an implied query surface. Josh expressed a similar
interpretation Slack
<https://apache-ossie.slack.com/archives/C0ADR49SBEK/p1790818302175719>, so
I don't think this interpretation is unique to me.

Allowing execute_query and other surfaces to run native-dialect queries
against a model would be a meaningful expansion of that concept. It would
turn dialects from optional expression representations into executable
query languages, even though the spec explicitly does not define shared
semantics across them and has not yet defined the shape or semantics of
such a query interface.

As far as I can tell, any native-dialect execution tool built today would
evaluate vendor expressions using vendor-specific semantics. Ossie could
standardize the tool shape, but not the meaning of the query or its result
across implementations. That looks much closer to a generic agent-to-engine
interface than an Ossie query interface.

So +1 to narrowing #529 to the model-resolved case. Since that case depends
on query interfaces Ossie has not landed yet, I think #529 should add the
caveat that the query interface shape isn't settled.

Separately, I'll start a thread on the scope of vendor-specific languages
so we can make explicit what registering a dialect does and does not imply
for execution.

Kind regards,
Chris

On Wed, Oct 7, 2026 at 4:59 AM Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:

> Hi all,
>
> I think Chris's question about how tightly execute_query ties to Ossie
> semantics decides where this should live.
>
> Ossie is an interchange spec for semantic models (at least today ;)). A
> running MCP server brings in execution, connections, auth and a fast-moving
> protocol surface. I would rather not take on in core, so my preference is
> for MCP to live in engines and catalogs, as I said earlier. The spec could
> make sense (like the REST API) but not the server implementation.
>
> The semantic question matters more than the location, though. I see two
> different things here:
> 1. The declarative queries (metrics, dimensions, filters) are resolved
> against the model. That's the way Aleksandr's ossie-clickhouse is working.
> The semantics come from the model, so two implementations should return the
> same answer for the same metric. That is a contract Ossie can legitimately
> define.
> 2. This is a generic agent to query engine interface. Ossie could not say
> whether two servers compute a metric the same way, so the spec would add
> nothing here.
>
> If we want anything shared, I would keep it small and non-normative. It
> would be a declarative query shape plus a tool schema for model discovery,
> published as a reference or in a separate repo, with no server code in this
> repository. Discovery tools can stay optional.
>
> Markus, would you be open to narrow PR #529 to the declarative case? I
> would also like to hear from other engine and catalog people on whether
> they would implement a shared shape.
>
> Regards
> JB
>
> On Wed, Oct 7, 2026 at 11:02 AM Markus Cozowicz <m...@cozowicz.at> wrote:
>
>> Hi Chris,
>>
>> Maybe being a bit pragmatic here, but it a) might take a while until all
>> vendors have conversions and b) it's most likely that each vendor might
>> have some unique features that are not part of the standard yet, thus
>> allowing all registered Ossie dialects seems like a good starting point to
>> me.
>>
>> From a goal perspective, if execute_query (or whatever we call it), makes
>> it into the Ossie standard and subsequently into the training corpus for
>> LLMs, we all don't have to teach the various agents how to do query
>> execution, as it would just be common knowledge (and hopefully comes with
>> all the benefits: less tokens, lower latency, higher accuracy,...)
>>
>> Markus
>> ________________________________
>> From: Chris Eubank via dev <de...@ossie.apache.org>
>> Sent: Wednesday, October 7, 2026 1:36 AM
>> To: dev@ossie.apache.org <de...@ossie.apache.org>
>> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>> live) [ossie]
>>
>> Hi Markus, Alex, JB,
>>
>> Thanks for opening #529 - it’s great to have something concrete to
>> discuss.
>>
>> I'm curious about the intended direction of execute_query. To what extent
>> is the contract meant to be tied to Ossie semantics? Is the goal to guide
>> clients toward portable Ossie query interfaces, or also to standardize the
>> execution of vendor-specific query languages compiled outside Ossie? #529
>> currently seems closer to the latter.  Its ties to Ossie also look fairly
>> loose - the model identity is optional, and the result schema/logical
>> types
>> appear to be the main semantic link. Is that intentionally the level of
>> coupling to Ossie?
>>
>> I’m asking because these seem to give pretty different portability
>> guarantees. If the caller sends an engine-native query, do two tools
>> consuming the same Ossie model each need to translate the semantic
>> definitions themselves? If so, where do we ensure they preserve the same
>> semantics?
>>
>> The implementations we have today seem to illustrate the distinction.
>> Alex’s ossie-clickhouse takes a query in Ossie’s in-flight Layer 3 terms
>> (metrics, dimensions, filters) and compiles it, whereas #529 accepts a
>> native query string. On binding, they seem aligned: the model carries no
>> connection, the server owns it, and the agent sees the exposed tools. How
>> do you see these two approaches fitting under a common contract?
>>
>> This also makes me wonder about scope. A vendor-agnostic execute_query
>> interface for agents seems useful, but if the query itself is
>> engine-native, I’m not yet sure what makes the contract specifically Ossie
>> rather than a more general agent-to-query-engine interface.
>>
>> JB, this may connect to your point about MCP looking closer to an
>> engine/catalog concern - curious how you see the boundary.
>>
>> Kind regards,
>> Chris
>>
>> On Tue, Oct 6, 2026 at 12:24 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>>
>> > Hi Alex,
>> >
>> > Good see we're following similar patterns.
>> >
>> > Input of execute_query: I'm not sure we should limit the input of
>> execute
>> > query to a specific layer or dialect. Some of our implementations
>> support
>> > 4-5 different query languages, depending on the targeted data engine.
>> >
>> > As for filter values: the canonical example: SELECT * FROM X WHERE
>> > CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as
>> > the filter value. In a simplified version, you can check if the results
>> are
>> > empty and the using backend logic figure what are alternative values you
>> > present to the agent.
>> >
>> > The below seems very similar to how we do things.
>> > - The model carries no connection.
>> > - The server starts with a model and a database connection.
>> > - The agent sees only the tools.
>> >
>> > I created a PR https://github.com/apache/ossie/pull/529 to get us
>> started
>> > on more details.
>> >
>> >
>> > Markus
>> > ________________________________
>> > From: Aleksandr Kosachev <a....@gmail.com>
>> > Sent: Tuesday, October 6, 2026 1:22 AM
>> > To: dev@ossie.apache.org <de...@ossie.apache.org>
>> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
>> > live) [ossie]
>> >
>> > Hi Markus, JB,
>> >
>> > One gap the thread has not named yet. The REST API spec proposed on
>> 09-23
>> > covers two things: managing models and binding a model to physical
>> > entities. It does not cover running a query and returning data. That is
>> > where execute_query starts. So an MCP spec next to the REST spec would
>> not
>> > duplicate it. It would fill the missing step.
>> >
>> > Markus asked how an agent that found a model knows how to run a query
>> > against the source.
>> > In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain,
>> it
>> > does not need to know:
>> > - The model carries no connection.
>> > - The server starts with a model and a database connection.
>> > - The agent sees only the tools.
>> >
>> > This reconciles the two positions in the thread. Ossie specifies the
>> tool
>> > contract. The engine or catalog that serves it owns the binding.
>> >
>> > One question needs an answer before a spec: what is the input of
>> > execute_query? PR #452 names two query interfaces:
>> > - Layer 2: SQL with MEASURE().
>> > - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>> >
>> > execute_query could accept either. ossie-clickhouse takes the Layer 3
>> form.
>> > When its planner refuses a query, the refusal names the cause.
>> > For example, "unknown metric 'x' (did you mean: ...)", or which metric
>> > would be counted once per joined row.
>> > The agent reads that and repairs the query.
>> > Markus listed four outputs: data, errors, suggestions, filter value
>> > alternatives.
>> > ossie-clickhouse returns data, errors and name suggestions, plus the
>> > generated SQL. It has no filter value alternatives yet.
>> > I would like to hear how those look on your side.
>> >
>> > On query_model and the harness dependency, ossie-clickhouse confirms it.
>> > Its overview tool returns datasets and metrics without fields.
>> > Agents then call a search tool and a describe-one-object tool before
>> every
>> > query. A harness with the file on disk, as Markus described, does not
>> need
>> > them. So discovery tools belong in the spec as optional, not dropped.
>> >
>> > Warm regards,
>> > Alex Kosachev
>> >
>> > [1] https://github.com/Dactopus/ossie-clickhouse
>> > <
>> >
>> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
>> > >
>> >
>> > On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>> >
>> > > I'm curious about the expectations here. If e.g. GitHub copilot
>> discovers
>> > > an Ossie semantic model, how would it know how to connect/execute a
>> query
>> > > against the data source (e.g. PBI semantic model, Databricks,
>> Snowflake,
>> > > ...)? Is the expectation that the harness would get additional
>> > instructions
>> > > anyway?
>> > >
>> > > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
>> > > implementation.
>> > > ________________________________
>> > > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
>> > > Sent: Monday, October 5, 2026 8:25 PM
>> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
>> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
>> should
>> > > live) [ossie]
>> > >
>> > > Hi
>> > >
>> > > It's not clear to me if we are discussing the specification or the
>> > > implementation. MCP implementation in Ossie: I'm not convinced; it
>> seems
>> > > more like an engine/catalog implementation.
>> > > The spec could make sense.
>> > >
>> > > But do we really need the MCP spec in Ossie for agent model discovery?
>> > I'm
>> > > not sure. That's actually one of the reasons for the REST API Spec too
>> > > (model discovery, versioning, etc).
>> > > We could have the MCP spec alongside the REST API Spec.
>> > >
>> > > Also, MCP is good to have but clearly not a requirement for the agent
>> > (API
>> > > or CLI are classic data providers for agent, including operations).
>> > >
>> > > Regards
>> > > JB
>> > >
>> > > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>> > >
>> > > > I'd like to provide a slightly different pitch for an MCP server
>> spec
>> > > (not
>> > > > necessarily the implementation). From my perspective, Ossies biggest
>> > > value
>> > > > prop is related to empowering agents to understand semantics and
>> > > > subsequently build queries to retrieve data.
>> > > > Most important is probably an execute_query MCP tool.
>> > > >
>> > > >
>> > > >   *
>> > > > Input: data source id, query
>> > > >   *
>> > > > Output
>> > > >      *
>> > > > Data
>> > > >      *
>> > > > Errors
>> > > >      *
>> > > > Suggestions
>> > > >      *
>> > > > Filter value alternatives
>> > > >      *
>> > > > ...
>> > > >
>> > > > This operates under the assumption that the agent is able to
>> > discover/see
>> > > > the Ossie semantic model (e.g. using
>> > > http://agenticresourcediscovery.org/).
>> > > > We already have a few variants of execute_query tool across our
>> > engines.
>> > > >
>> > > > I'm not sure about query_model (yet), at least for now we see a
>> > > dependency
>> > > > on the agent harness. Some harness can only digest a small view of
>> the
>> > > > semantic model and thus need tools to progressively discover, others
>> > > (like
>> > > > e.g. GitHub copilot cli), don't need any of that, as they'll have
>> full
>> > > > filesystem access, download/query local.
>> > > >
>> > > > ________________________________
>> > > > From: jbonofre (via GitHub) <gi...@apache.org>
>> > > > Sent: Monday, October 5, 2026 8:01 AM
>> > > > To: dev@ossie.apache.org <de...@ossie.apache.org>
>> > > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
>> should
>> > > > live) [ossie]
>> > > >
>> > > >
>> > > > GitHub user jbonofre added a comment to the discussion: Proposal: an
>> > MCP
>> > > > server for Ossie (and where it should live)
>> > > >
>> > > > I'm not sure we need MCP server at all in Ossie. That could be
>> > something
>> > > > catalogs can do.
>> > > >
>> > > > I would rather not consider MCP service directly in Ossie for now (I
>> > > don't
>> > > > see the value).
>> > > >
>> > > > GitHub link:
>> > > >
>> > >
>> >
>> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>> > > >
>> > > > ----
>> > > > This is an automatically sent email for dev@ossie.apache.org.
>> > > > To unsubscribe, please send an email to:
>> > > dev-unsubscribe@ossie.apache.org
>> > > >
>> > > >
>> > >
>> >
>>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
@Jean-Baptiste Onofré<ma...@nanthrax.net> it should already be narrowed to declarative case.

Markus
________________________________
From: Jean-Baptiste Onofré <jb...@nanthrax.net>
Sent: Tuesday, October 13, 2026 10:10 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Cc: willpugh@apache.org <wi...@apache.org>; christian.eubank@databricks.com <ch...@databricks.com>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Hi Alex

Thanks you so much! The questions are super super interesting and it helps
to "fuel" the discussion.

I have a few thoughts.

First, imho, a second engine is the next "real" test :) Your cases, data
and captured answers are ready to be reused imho. So it would be great to
see another engine run them. Where the answers differ, I think PR #246 is
underspecified, and those are the cases to fix in the spec. I will find an
engine candidate to do a "reply scenario".

About T-013 vs 6.6 row 4, I think we have a kind of disagreement between
these parts ;) I read 6.6 row 4 as correct, since two one-row totals side
by side are well-defined, so 6.8.2 and T-013 in #237 should not expect
E3013 for that query. Your executor already does this. I propose to settle
it on PR #246 and update T-013 to match.

I think your question 5 is the most important open point. If an engine
can't compute a query that #246 answers, it needs a shared code with a
shared list of named capability restrictions. Otherwise every engine will
invent its own, and the codes stop meaning the same thing across engines. I
think what you proposed in #246 is a good starting point.

Your 80 sessions suggest agents repair queries from the message and the
did-you-mean suggestion about as well without codes. I read that as support
for keeping the shared code list small and stable. The codes matter mainly
for programs around the agent and for consistent refusals across engines.
It also makes me want to be careful about maintaining 28 codes identically
across engines.

Related, I agree that requiring adapter SQL to run in DuckDB excludes
engines with their own dialect. I would prefer the suite compare results
against expected rows, so a Clickhouse engine can take part. We can discuss
about that (cases listing) on the mailing list.

I'm waiting for Markus on #529. My preference is to narrow it to the
declarative case, so the #246 code can be error.code and replacement_query
can be a Layer 3 object.

Thanks again for this great work!

Regards
JB

On Tue, Oct 13, 2026 at 3:04 AM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi all,
>
> Following up on JB's point that the refusals tested the shape, not the
> semantics [1]. The executor now answers two of the three kinds of queries
> it used to refuse, and part of the third. The answers match expected rows
> derived from #246 line by line. What is still missing is a second engine
> running the same cases.
>
> What changed in dactopus-ossie-clickhouse (formerly ossie-clickhouse):
> - paths of several N:1 relationships are followed, and a second path is
> refused with E_AMBIGUOUS_PATH;
> - measures over different datasets are merged with a NULL-safe FULL OUTER
> JOIN on the dimensions, or a CROSS JOIN of totals when there are none;
> - each aggregate of an arithmetic metric is computed over its own
> dataset's rows (in a draft PR, since it changes released numbers);
> pre-aggregation (§6.7) is still refused.
>
> How the answers were checked:
> - the expected rows cite a line of #246 at cc0d070 for each step of the
> plan;
> - on synthetic data, four wrong plans give different numbers, for example
> a FULL OUTER JOIN by plain equality;
> - eleven cases in the layout of the compliance suite (#237) match the
> expected rows;
> - on the model with two facts, E3013, E_NO_PATH and the three Having codes
> come where #246 puts them, and the Having answers match the expected rows;
> - in ten agent sessions on that model, E3013 and E_NO_PATH each led to a
> right answer or a right decline in the next step.
>
> The expected rows are my reading of #246. A second engine would most
> likely differ where #246 is still open. I asked about those points on #246
> [2]: totals of two unrelated facts, a path that changes direction, a second
> and longer path, and an aggregate over a coarser dataset.
>
> The first one now has two readings in the project's own texts. JB read
> §6.6 row 4 as one row of totals, and the executor answers that way. T-013
> in #237 expects E3013 for the same query.
>
> On question 4, the executor now refuses a query for item and store
> dimensions alone, as JB would, with E_NO_PATH. On question 5, #529 already
> asks bindings to advertise capability restrictions, but there is no shared
> list of names. I proposed one on #246 [2].
>
> The cases are ready for a second engine. The #237 harness, though, runs
> the adapter's SQL in DuckDB, and an engine with its own SQL dialect cannot
> take part that way. I asked about that and offered the cases on #237 [3].
> The data, cases and captured answers are in [4].
>
> Warm regards,
> Alex
>
> [1] https://lists.apache.org/thread/9007r91vf33y04grtk8ynv58nqc4oyco,
> JB's message of 10-08
> [2] https://github.com/apache/ossie/pull/246, questions on #246
> [3] https://github.com/apache/ossie/pull/237#pullrequestreview-5497883375,
> questions on #237
> [4] https://gist.github.com/ekut/25fe42cb6766d9a75fe0083281ebdb52, data
> and cases
>
>
>
> > On 8 Oct 2026, at 11.44, Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:
> >
> > Hi Alex,
> >
> > Thanks, the engine perspective and feedback is important here.
> >
> > A few takeaways from my side:
> >
> > 1. The declarative shape works as a shared input. You implemented it
> > quickly, outside this repository, and the agents has no trouble with it.
> > For me that supports keeping the shared part small (a query shape plus
> > error codes) and the server in the engines.
> >
> > 2. Your two questions for #529 both come from it taking a native query
> > string. If #529 is narrowed to the declarative case, query and
> > replacement_query become Layer 3 objects and #246  code can simply be
> > error.code. If it stays generic, the #246 code would have to set under
> > extensions. Markus, what do you think?
> >
> > 3. What the executor refuses today (multi-stop paths, measures over
> > different datasets, metrics over a joined dataset's rows) is the part
> that
> > makes two engines return the same number. Refusing is the right behavior
> > imho, much better than a wrong answer. But it means we have tested the
> > shape so far, not the semantics.
> >
> > On the #246 questions, my reading:
> > - Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by
> side
> > are correct numbers, so I don't see why 6.8.2 should refuse.
> > - Question 4: I would refuse. In TPC-DS, item and store meet through both
> > store_sales and store_returns, so the result depends on a fact the query
> > does not name.
> > - Questiion 5: this is the important one. #246 only has a capability
> > opt-out for M:N. I think we need a general one, or each engine will pick
> > its own code.
> >
> > Could we post the #246 questions as review comments on the PR? It will be
> > easier for Will (and others) to resolve them there than in this thread,
> and
> > the list keeps the summary/decision.
> >
> > As a next step, a model with two facts would exercise E3013, E_NO_PATH
> and
> > Having. And if a second engine implements the same shape, we can compare
> > answers and not just shapes.
> >
> > Regards
> > JB
> >
> > On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a.yakutskiy@gmail.com
> >
> > wrote:
> >
> >> Hi all,
> >>
> >> Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
> >> now takes the aggregation query from #246 §5.1.1 as a JSON object.
> Refusals
> >> carry #246 codes in the #529 error envelope. Implementing it raised a
> few
> >> questions about both drafts. The full report, with line references, is
> at
> >> [2]. Will, I'm copying you because most of the questions are about #246.
> >>
> >> The three things I promised to report:
> >> - The executor raises 10 of the 28 #246 codes as specified, and
> >> E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
> >> - Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY
> when
> >> the query is valid but the executor cannot compute it, QUERY_INVALID for
> >> parse errors.
> >> - Three kinds of queries that #246 expects answered get
> UNSUPPORTED_QUERY:
> >> multi-step relationship paths, measures over different datasets in one
> >> query, and metrics over a joined dataset's rows.
> >>
> >> Questions for #246:
> >> 1. The object form of Order By has no key for NULLS FIRST or LAST,
> although
> >> §5.1 lets an entry specify it. I added a nulls key as an extension.
> >> 2. Does E3013 apply within one measure, as the §6.2 table says, or
> within
> >> one query, as Appendix A says?
> >> 3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is
> empty,
> >> since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
> >> totals with a CROSS JOIN, which gives one row. Which rule applies?
> >> 4. Can a path between two dimension datasets run through a fact? §6.9
> >> builds paths from N:1 and 1:1 edges only.
> >> 5. Which code should an engine use for a query that #246 answers but the
> >> engine does not? Semantic 5 lets engines differ on M:N edge cases, and
> >> E3011 opts out of M:N as a whole. The executor also refuses, for
> example,
> >> two measures over different datasets in one query, and finds no #246
> code
> >> for that.
> >>
> >> Two questions for #529, when the input is a Layer 3 object. Should the
> #246
> >> code replace the #529 code in error.code, or sit next to it under
> >> extensions? And what should a replacement query look like, given that
> query
> >> and replacement_query are strings while a Layer 3 query is an object?
> >>
> >> I also ran 80 agent sessions: 20 analyst questions, two models, the new
> >> input and the previous one. Agents made no errors in the shape.
> >>
> >> Yesterday I wrote that the codes matter because the agent repairs its
> query
> >> from the refusal. The run does not show that: agents repaired from the
> >> message and the did-you-mean suggestion, and did about as well without
> >> codes. Codes may matter more to programs around the agent, and for a
> >> refusal to mean the same on every engine. Refusals such as E3013 and
> >> E_NO_PATH never came up, because the test model has one central dataset.
> >>
> >> Warm regards,
> >> Alex
> >>
> >> [1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
> >> <
> >>
> https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E
> >>> ,
> >> my message of 10-07
> >> [2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
> >> <
> >>
> https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E
> >>> ,
> >> the full report
> >>
> >> On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <
> a.yakutskiy@gmail.com>
> >> wrote:
> >>
> >> Hi all,
> >>
> >> On JB's question about engines: I'd like to offer ossie-clickhouse [1]
> as a
> >> test bed for a shared shape. I'm starting now and will report to this
> >> thread as results come in.
> >> [...]
> >>
>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Jean-Baptiste Onofré <jb...@nanthrax.net>.
Hi Alex

Thanks you so much! The questions are super super interesting and it helps
to "fuel" the discussion.

I have a few thoughts.

First, imho, a second engine is the next "real" test :) Your cases, data
and captured answers are ready to be reused imho. So it would be great to
see another engine run them. Where the answers differ, I think PR #246 is
underspecified, and those are the cases to fix in the spec. I will find an
engine candidate to do a "reply scenario".

About T-013 vs 6.6 row 4, I think we have a kind of disagreement between
these parts ;) I read 6.6 row 4 as correct, since two one-row totals side
by side are well-defined, so 6.8.2 and T-013 in #237 should not expect
E3013 for that query. Your executor already does this. I propose to settle
it on PR #246 and update T-013 to match.

I think your question 5 is the most important open point. If an engine
can't compute a query that #246 answers, it needs a shared code with a
shared list of named capability restrictions. Otherwise every engine will
invent its own, and the codes stop meaning the same thing across engines. I
think what you proposed in #246 is a good starting point.

Your 80 sessions suggest agents repair queries from the message and the
did-you-mean suggestion about as well without codes. I read that as support
for keeping the shared code list small and stable. The codes matter mainly
for programs around the agent and for consistent refusals across engines.
It also makes me want to be careful about maintaining 28 codes identically
across engines.

Related, I agree that requiring adapter SQL to run in DuckDB excludes
engines with their own dialect. I would prefer the suite compare results
against expected rows, so a Clickhouse engine can take part. We can discuss
about that (cases listing) on the mailing list.

I'm waiting for Markus on #529. My preference is to narrow it to the
declarative case, so the #246 code can be error.code and replacement_query
can be a Layer 3 object.

Thanks again for this great work!

Regards
JB

On Tue, Oct 13, 2026 at 3:04 AM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi all,
>
> Following up on JB's point that the refusals tested the shape, not the
> semantics [1]. The executor now answers two of the three kinds of queries
> it used to refuse, and part of the third. The answers match expected rows
> derived from #246 line by line. What is still missing is a second engine
> running the same cases.
>
> What changed in dactopus-ossie-clickhouse (formerly ossie-clickhouse):
> - paths of several N:1 relationships are followed, and a second path is
> refused with E_AMBIGUOUS_PATH;
> - measures over different datasets are merged with a NULL-safe FULL OUTER
> JOIN on the dimensions, or a CROSS JOIN of totals when there are none;
> - each aggregate of an arithmetic metric is computed over its own
> dataset's rows (in a draft PR, since it changes released numbers);
> pre-aggregation (§6.7) is still refused.
>
> How the answers were checked:
> - the expected rows cite a line of #246 at cc0d070 for each step of the
> plan;
> - on synthetic data, four wrong plans give different numbers, for example
> a FULL OUTER JOIN by plain equality;
> - eleven cases in the layout of the compliance suite (#237) match the
> expected rows;
> - on the model with two facts, E3013, E_NO_PATH and the three Having codes
> come where #246 puts them, and the Having answers match the expected rows;
> - in ten agent sessions on that model, E3013 and E_NO_PATH each led to a
> right answer or a right decline in the next step.
>
> The expected rows are my reading of #246. A second engine would most
> likely differ where #246 is still open. I asked about those points on #246
> [2]: totals of two unrelated facts, a path that changes direction, a second
> and longer path, and an aggregate over a coarser dataset.
>
> The first one now has two readings in the project's own texts. JB read
> §6.6 row 4 as one row of totals, and the executor answers that way. T-013
> in #237 expects E3013 for the same query.
>
> On question 4, the executor now refuses a query for item and store
> dimensions alone, as JB would, with E_NO_PATH. On question 5, #529 already
> asks bindings to advertise capability restrictions, but there is no shared
> list of names. I proposed one on #246 [2].
>
> The cases are ready for a second engine. The #237 harness, though, runs
> the adapter's SQL in DuckDB, and an engine with its own SQL dialect cannot
> take part that way. I asked about that and offered the cases on #237 [3].
> The data, cases and captured answers are in [4].
>
> Warm regards,
> Alex
>
> [1] https://lists.apache.org/thread/9007r91vf33y04grtk8ynv58nqc4oyco,
> JB's message of 10-08
> [2] https://github.com/apache/ossie/pull/246, questions on #246
> [3] https://github.com/apache/ossie/pull/237#pullrequestreview-5497883375,
> questions on #237
> [4] https://gist.github.com/ekut/25fe42cb6766d9a75fe0083281ebdb52, data
> and cases
>
>
>
> > On 8 Oct 2026, at 11.44, Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:
> >
> > Hi Alex,
> >
> > Thanks, the engine perspective and feedback is important here.
> >
> > A few takeaways from my side:
> >
> > 1. The declarative shape works as a shared input. You implemented it
> > quickly, outside this repository, and the agents has no trouble with it.
> > For me that supports keeping the shared part small (a query shape plus
> > error codes) and the server in the engines.
> >
> > 2. Your two questions for #529 both come from it taking a native query
> > string. If #529 is narrowed to the declarative case, query and
> > replacement_query become Layer 3 objects and #246  code can simply be
> > error.code. If it stays generic, the #246 code would have to set under
> > extensions. Markus, what do you think?
> >
> > 3. What the executor refuses today (multi-stop paths, measures over
> > different datasets, metrics over a joined dataset's rows) is the part
> that
> > makes two engines return the same number. Refusing is the right behavior
> > imho, much better than a wrong answer. But it means we have tested the
> > shape so far, not the semantics.
> >
> > On the #246 questions, my reading:
> > - Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by
> side
> > are correct numbers, so I don't see why 6.8.2 should refuse.
> > - Question 4: I would refuse. In TPC-DS, item and store meet through both
> > store_sales and store_returns, so the result depends on a fact the query
> > does not name.
> > - Questiion 5: this is the important one. #246 only has a capability
> > opt-out for M:N. I think we need a general one, or each engine will pick
> > its own code.
> >
> > Could we post the #246 questions as review comments on the PR? It will be
> > easier for Will (and others) to resolve them there than in this thread,
> and
> > the list keeps the summary/decision.
> >
> > As a next step, a model with two facts would exercise E3013, E_NO_PATH
> and
> > Having. And if a second engine implements the same shape, we can compare
> > answers and not just shapes.
> >
> > Regards
> > JB
> >
> > On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a.yakutskiy@gmail.com
> >
> > wrote:
> >
> >> Hi all,
> >>
> >> Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
> >> now takes the aggregation query from #246 §5.1.1 as a JSON object.
> Refusals
> >> carry #246 codes in the #529 error envelope. Implementing it raised a
> few
> >> questions about both drafts. The full report, with line references, is
> at
> >> [2]. Will, I'm copying you because most of the questions are about #246.
> >>
> >> The three things I promised to report:
> >> - The executor raises 10 of the 28 #246 codes as specified, and
> >> E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
> >> - Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY
> when
> >> the query is valid but the executor cannot compute it, QUERY_INVALID for
> >> parse errors.
> >> - Three kinds of queries that #246 expects answered get
> UNSUPPORTED_QUERY:
> >> multi-step relationship paths, measures over different datasets in one
> >> query, and metrics over a joined dataset's rows.
> >>
> >> Questions for #246:
> >> 1. The object form of Order By has no key for NULLS FIRST or LAST,
> although
> >> §5.1 lets an entry specify it. I added a nulls key as an extension.
> >> 2. Does E3013 apply within one measure, as the §6.2 table says, or
> within
> >> one query, as Appendix A says?
> >> 3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is
> empty,
> >> since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
> >> totals with a CROSS JOIN, which gives one row. Which rule applies?
> >> 4. Can a path between two dimension datasets run through a fact? §6.9
> >> builds paths from N:1 and 1:1 edges only.
> >> 5. Which code should an engine use for a query that #246 answers but the
> >> engine does not? Semantic 5 lets engines differ on M:N edge cases, and
> >> E3011 opts out of M:N as a whole. The executor also refuses, for
> example,
> >> two measures over different datasets in one query, and finds no #246
> code
> >> for that.
> >>
> >> Two questions for #529, when the input is a Layer 3 object. Should the
> #246
> >> code replace the #529 code in error.code, or sit next to it under
> >> extensions? And what should a replacement query look like, given that
> query
> >> and replacement_query are strings while a Layer 3 query is an object?
> >>
> >> I also ran 80 agent sessions: 20 analyst questions, two models, the new
> >> input and the previous one. Agents made no errors in the shape.
> >>
> >> Yesterday I wrote that the codes matter because the agent repairs its
> query
> >> from the refusal. The run does not show that: agents repaired from the
> >> message and the did-you-mean suggestion, and did about as well without
> >> codes. Codes may matter more to programs around the agent, and for a
> >> refusal to mean the same on every engine. Refusals such as E3013 and
> >> E_NO_PATH never came up, because the test model has one central dataset.
> >>
> >> Warm regards,
> >> Alex
> >>
> >> [1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
> >> <
> >>
> https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E
> >>> ,
> >> my message of 10-07
> >> [2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
> >> <
> >>
> https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E
> >>> ,
> >> the full report
> >>
> >> On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <
> a.yakutskiy@gmail.com>
> >> wrote:
> >>
> >> Hi all,
> >>
> >> On JB's question about engines: I'd like to offer ossie-clickhouse [1]
> as a
> >> test bed for a shared shape. I'm starting now and will report to this
> >> thread as results come in.
> >> [...]
> >>
>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi all,

Following up on JB's point that the refusals tested the shape, not the semantics [1]. The executor now answers two of the three kinds of queries it used to refuse, and part of the third. The answers match expected rows derived from #246 line by line. What is still missing is a second engine running the same cases.

What changed in dactopus-ossie-clickhouse (formerly ossie-clickhouse):
- paths of several N:1 relationships are followed, and a second path is refused with E_AMBIGUOUS_PATH;
- measures over different datasets are merged with a NULL-safe FULL OUTER JOIN on the dimensions, or a CROSS JOIN of totals when there are none;
- each aggregate of an arithmetic metric is computed over its own dataset's rows (in a draft PR, since it changes released numbers); pre-aggregation (§6.7) is still refused.

How the answers were checked:
- the expected rows cite a line of #246 at cc0d070 for each step of the plan;
- on synthetic data, four wrong plans give different numbers, for example a FULL OUTER JOIN by plain equality;
- eleven cases in the layout of the compliance suite (#237) match the expected rows;
- on the model with two facts, E3013, E_NO_PATH and the three Having codes come where #246 puts them, and the Having answers match the expected rows;
- in ten agent sessions on that model, E3013 and E_NO_PATH each led to a right answer or a right decline in the next step.

The expected rows are my reading of #246. A second engine would most likely differ where #246 is still open. I asked about those points on #246 [2]: totals of two unrelated facts, a path that changes direction, a second and longer path, and an aggregate over a coarser dataset.

The first one now has two readings in the project's own texts. JB read §6.6 row 4 as one row of totals, and the executor answers that way. T-013 in #237 expects E3013 for the same query.

On question 4, the executor now refuses a query for item and store dimensions alone, as JB would, with E_NO_PATH. On question 5, #529 already asks bindings to advertise capability restrictions, but there is no shared list of names. I proposed one on #246 [2].

The cases are ready for a second engine. The #237 harness, though, runs the adapter's SQL in DuckDB, and an engine with its own SQL dialect cannot take part that way. I asked about that and offered the cases on #237 [3]. The data, cases and captured answers are in [4].

Warm regards,
Alex

[1] https://lists.apache.org/thread/9007r91vf33y04grtk8ynv58nqc4oyco, JB's message of 10-08
[2] https://github.com/apache/ossie/pull/246, questions on #246
[3] https://github.com/apache/ossie/pull/237#pullrequestreview-5497883375, questions on #237
[4] https://gist.github.com/ekut/25fe42cb6766d9a75fe0083281ebdb52, data and cases



> On 8 Oct 2026, at 11.44, Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:
> 
> Hi Alex,
> 
> Thanks, the engine perspective and feedback is important here.
> 
> A few takeaways from my side:
> 
> 1. The declarative shape works as a shared input. You implemented it
> quickly, outside this repository, and the agents has no trouble with it.
> For me that supports keeping the shared part small (a query shape plus
> error codes) and the server in the engines.
> 
> 2. Your two questions for #529 both come from it taking a native query
> string. If #529 is narrowed to the declarative case, query and
> replacement_query become Layer 3 objects and #246  code can simply be
> error.code. If it stays generic, the #246 code would have to set under
> extensions. Markus, what do you think?
> 
> 3. What the executor refuses today (multi-stop paths, measures over
> different datasets, metrics over a joined dataset's rows) is the part that
> makes two engines return the same number. Refusing is the right behavior
> imho, much better than a wrong answer. But it means we have tested the
> shape so far, not the semantics.
> 
> On the #246 questions, my reading:
> - Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by side
> are correct numbers, so I don't see why 6.8.2 should refuse.
> - Question 4: I would refuse. In TPC-DS, item and store meet through both
> store_sales and store_returns, so the result depends on a fact the query
> does not name.
> - Questiion 5: this is the important one. #246 only has a capability
> opt-out for M:N. I think we need a general one, or each engine will pick
> its own code.
> 
> Could we post the #246 questions as review comments on the PR? It will be
> easier for Will (and others) to resolve them there than in this thread, and
> the list keeps the summary/decision.
> 
> As a next step, a model with two facts would exercise E3013, E_NO_PATH and
> Having. And if a second engine implements the same shape, we can compare
> answers and not just shapes.
> 
> Regards
> JB
> 
> On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a....@gmail.com>
> wrote:
> 
>> Hi all,
>> 
>> Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
>> now takes the aggregation query from #246 §5.1.1 as a JSON object. Refusals
>> carry #246 codes in the #529 error envelope. Implementing it raised a few
>> questions about both drafts. The full report, with line references, is at
>> [2]. Will, I'm copying you because most of the questions are about #246.
>> 
>> The three things I promised to report:
>> - The executor raises 10 of the 28 #246 codes as specified, and
>> E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
>> - Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY when
>> the query is valid but the executor cannot compute it, QUERY_INVALID for
>> parse errors.
>> - Three kinds of queries that #246 expects answered get UNSUPPORTED_QUERY:
>> multi-step relationship paths, measures over different datasets in one
>> query, and metrics over a joined dataset's rows.
>> 
>> Questions for #246:
>> 1. The object form of Order By has no key for NULLS FIRST or LAST, although
>> §5.1 lets an entry specify it. I added a nulls key as an extension.
>> 2. Does E3013 apply within one measure, as the §6.2 table says, or within
>> one query, as Appendix A says?
>> 3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is empty,
>> since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
>> totals with a CROSS JOIN, which gives one row. Which rule applies?
>> 4. Can a path between two dimension datasets run through a fact? §6.9
>> builds paths from N:1 and 1:1 edges only.
>> 5. Which code should an engine use for a query that #246 answers but the
>> engine does not? Semantic 5 lets engines differ on M:N edge cases, and
>> E3011 opts out of M:N as a whole. The executor also refuses, for example,
>> two measures over different datasets in one query, and finds no #246 code
>> for that.
>> 
>> Two questions for #529, when the input is a Layer 3 object. Should the #246
>> code replace the #529 code in error.code, or sit next to it under
>> extensions? And what should a replacement query look like, given that query
>> and replacement_query are strings while a Layer 3 query is an object?
>> 
>> I also ran 80 agent sessions: 20 analyst questions, two models, the new
>> input and the previous one. Agents made no errors in the shape.
>> 
>> Yesterday I wrote that the codes matter because the agent repairs its query
>> from the refusal. The run does not show that: agents repaired from the
>> message and the did-you-mean suggestion, and did about as well without
>> codes. Codes may matter more to programs around the agent, and for a
>> refusal to mean the same on every engine. Refusals such as E3013 and
>> E_NO_PATH never came up, because the test model has one central dataset.
>> 
>> Warm regards,
>> Alex
>> 
>> [1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
>> <
>> https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E
>>> ,
>> my message of 10-07
>> [2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
>> <
>> https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E
>>> ,
>> the full report
>> 
>> On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <a....@gmail.com>
>> wrote:
>> 
>> Hi all,
>> 
>> On JB's question about engines: I'd like to offer ossie-clickhouse [1] as a
>> test bed for a shared shape. I'm starting now and will report to this
>> thread as results come in.
>> [...]
>> 


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
Regarding declarative case: works for me - I updated the PR.

From an adoption perspective, allowing other layers would make it easier.

Best,
Markus
________________________________
From: Jean-Baptiste Onofré <jb...@nanthrax.net>
Sent: Thursday, October 8, 2026 5:44 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Cc: willpugh@apache.org <wi...@apache.org>; christian.eubank@databricks.com <ch...@databricks.com>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Hi Alex,

Thanks, the engine perspective and feedback is important here.

A few takeaways from my side:

1. The declarative shape works as a shared input. You implemented it
quickly, outside this repository, and the agents has no trouble with it.
For me that supports keeping the shared part small (a query shape plus
error codes) and the server in the engines.

2. Your two questions for #529 both come from it taking a native query
string. If #529 is narrowed to the declarative case, query and
replacement_query become Layer 3 objects and #246  code can simply be
error.code. If it stays generic, the #246 code would have to set under
extensions. Markus, what do you think?

3. What the executor refuses today (multi-stop paths, measures over
different datasets, metrics over a joined dataset's rows) is the part that
makes two engines return the same number. Refusing is the right behavior
imho, much better than a wrong answer. But it means we have tested the
shape so far, not the semantics.

On the #246 questions, my reading:
- Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by side
are correct numbers, so I don't see why 6.8.2 should refuse.
- Question 4: I would refuse. In TPC-DS, item and store meet through both
store_sales and store_returns, so the result depends on a fact the query
does not name.
- Questiion 5: this is the important one. #246 only has a capability
opt-out for M:N. I think we need a general one, or each engine will pick
its own code.

Could we post the #246 questions as review comments on the PR? It will be
easier for Will (and others) to resolve them there than in this thread, and
the list keeps the summary/decision.

As a next step, a model with two facts would exercise E3013, E_NO_PATH and
Having. And if a second engine implements the same shape, we can compare
answers and not just shapes.

Regards
JB

On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi all,
>
> Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
> now takes the aggregation query from #246 §5.1.1 as a JSON object. Refusals
> carry #246 codes in the #529 error envelope. Implementing it raised a few
> questions about both drafts. The full report, with line references, is at
> [2]. Will, I'm copying you because most of the questions are about #246.
>
> The three things I promised to report:
> - The executor raises 10 of the 28 #246 codes as specified, and
> E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
> - Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY when
> the query is valid but the executor cannot compute it, QUERY_INVALID for
> parse errors.
> - Three kinds of queries that #246 expects answered get UNSUPPORTED_QUERY:
> multi-step relationship paths, measures over different datasets in one
> query, and metrics over a joined dataset's rows.
>
> Questions for #246:
> 1. The object form of Order By has no key for NULLS FIRST or LAST, although
> §5.1 lets an entry specify it. I added a nulls key as an extension.
> 2. Does E3013 apply within one measure, as the §6.2 table says, or within
> one query, as Appendix A says?
> 3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is empty,
> since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
> totals with a CROSS JOIN, which gives one row. Which rule applies?
> 4. Can a path between two dimension datasets run through a fact? §6.9
> builds paths from N:1 and 1:1 edges only.
> 5. Which code should an engine use for a query that #246 answers but the
> engine does not? Semantic 5 lets engines differ on M:N edge cases, and
> E3011 opts out of M:N as a whole. The executor also refuses, for example,
> two measures over different datasets in one query, and finds no #246 code
> for that.
>
> Two questions for #529, when the input is a Layer 3 object. Should the #246
> code replace the #529 code in error.code, or sit next to it under
> extensions? And what should a replacement query look like, given that query
> and replacement_query are strings while a Layer 3 query is an object?
>
> I also ran 80 agent sessions: 20 analyst questions, two models, the new
> input and the previous one. Agents made no errors in the shape.
>
> Yesterday I wrote that the codes matter because the agent repairs its query
> from the refusal. The run does not show that: agents repaired from the
> message and the did-you-mean suggestion, and did about as well without
> codes. Codes may matter more to programs around the agent, and for a
> refusal to mean the same on every engine. Refusals such as E3013 and
> E_NO_PATH never came up, because the test model has one central dataset.
>
> Warm regards,
> Alex
>
> [1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
> <
> https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E
> >,
> my message of 10-07
> [2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
> <
> https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E
> >,
> the full report
>
> On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <a....@gmail.com>
> wrote:
>
> Hi all,
>
> On JB's question about engines: I'd like to offer ossie-clickhouse [1] as a
> test bed for a shared shape. I'm starting now and will report to this
> thread as results come in.
> [...]
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi JB,

Thanks. The five #246 questions are now review comments on the PR, one per line, with your readings on questions 3 to 5 noted [1].

On the next step, I agree: a model with two facts is what will exercise E3013, E_NO_PATH and Having. 
I'll run it on the executor next week and report back here.

Warm regards,
Alex

[1] https://github.com/apache/ossie/pull/246#pullrequestreview-5452963210, the review with the five questions



> On 8 Oct 2026, at 11.44, Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:
> 
> Hi Alex,
> 
> Thanks, the engine perspective and feedback is important here.
> 
> A few takeaways from my side:
> 
> 1. The declarative shape works as a shared input. You implemented it
> quickly, outside this repository, and the agents has no trouble with it.
> For me that supports keeping the shared part small (a query shape plus
> error codes) and the server in the engines.
> 
> 2. Your two questions for #529 both come from it taking a native query
> string. If #529 is narrowed to the declarative case, query and
> replacement_query become Layer 3 objects and #246  code can simply be
> error.code. If it stays generic, the #246 code would have to set under
> extensions. Markus, what do you think?
> 
> 3. What the executor refuses today (multi-stop paths, measures over
> different datasets, metrics over a joined dataset's rows) is the part that
> makes two engines return the same number. Refusing is the right behavior
> imho, much better than a wrong answer. But it means we have tested the
> shape so far, not the semantics.
> 
> On the #246 questions, my reading:
> - Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by side
> are correct numbers, so I don't see why 6.8.2 should refuse.
> - Question 4: I would refuse. In TPC-DS, item and store meet through both
> store_sales and store_returns, so the result depends on a fact the query
> does not name.
> - Questiion 5: this is the important one. #246 only has a capability
> opt-out for M:N. I think we need a general one, or each engine will pick
> its own code.
> 
> Could we post the #246 questions as review comments on the PR? It will be
> easier for Will (and others) to resolve them there than in this thread, and
> the list keeps the summary/decision.
> 
> As a next step, a model with two facts would exercise E3013, E_NO_PATH and
> Having. And if a second engine implements the same shape, we can compare
> answers and not just shapes.
> 
> Regards
> JB
> 
> On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a.yakutskiy@gmail.com <ma...@gmail.com>>
> wrote:
> 


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Jean-Baptiste Onofré <jb...@nanthrax.net>.
Hi Alex,

Thanks, the engine perspective and feedback is important here.

A few takeaways from my side:

1. The declarative shape works as a shared input. You implemented it
quickly, outside this repository, and the agents has no trouble with it.
For me that supports keeping the shared part small (a query shape plus
error codes) and the server in the engines.

2. Your two questions for #529 both come from it taking a native query
string. If #529 is narrowed to the declarative case, query and
replacement_query become Layer 3 objects and #246  code can simply be
error.code. If it stays generic, the #246 code would have to set under
extensions. Markus, what do you think?

3. What the executor refuses today (multi-stop paths, measures over
different datasets, metrics over a joined dataset's rows) is the part that
makes two engines return the same number. Refusing is the right behavior
imho, much better than a wrong answer. But it means we have tested the
shape so far, not the semantics.

On the #246 questions, my reading:
- Questiion 3: 6.6 row 4 looks right to me. Two one-row totals side by side
are correct numbers, so I don't see why 6.8.2 should refuse.
- Question 4: I would refuse. In TPC-DS, item and store meet through both
store_sales and store_returns, so the result depends on a fact the query
does not name.
- Questiion 5: this is the important one. #246 only has a capability
opt-out for M:N. I think we need a general one, or each engine will pick
its own code.

Could we post the #246 questions as review comments on the PR? It will be
easier for Will (and others) to resolve them there than in this thread, and
the list keeps the summary/decision.

As a next step, a model with two facts would exercise E3013, E_NO_PATH and
Having. And if a second engine implements the same shape, we can compare
answers and not just shapes.

Regards
JB

On Thu, Oct 8, 2026 at 3:21 AM Aleksandr Kosachev <a....@gmail.com>
wrote:

> Hi all,
>
> Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
> now takes the aggregation query from #246 §5.1.1 as a JSON object. Refusals
> carry #246 codes in the #529 error envelope. Implementing it raised a few
> questions about both drafts. The full report, with line references, is at
> [2]. Will, I'm copying you because most of the questions are about #246.
>
> The three things I promised to report:
> - The executor raises 10 of the 28 #246 codes as specified, and
> E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
> - Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY when
> the query is valid but the executor cannot compute it, QUERY_INVALID for
> parse errors.
> - Three kinds of queries that #246 expects answered get UNSUPPORTED_QUERY:
> multi-step relationship paths, measures over different datasets in one
> query, and metrics over a joined dataset's rows.
>
> Questions for #246:
> 1. The object form of Order By has no key for NULLS FIRST or LAST, although
> §5.1 lets an entry specify it. I added a nulls key as an extension.
> 2. Does E3013 apply within one measure, as the §6.2 table says, or within
> one query, as Appendix A says?
> 3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is empty,
> since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
> totals with a CROSS JOIN, which gives one row. Which rule applies?
> 4. Can a path between two dimension datasets run through a fact? §6.9
> builds paths from N:1 and 1:1 edges only.
> 5. Which code should an engine use for a query that #246 answers but the
> engine does not? Semantic 5 lets engines differ on M:N edge cases, and
> E3011 opts out of M:N as a whole. The executor also refuses, for example,
> two measures over different datasets in one query, and finds no #246 code
> for that.
>
> Two questions for #529, when the input is a Layer 3 object. Should the #246
> code replace the #529 code in error.code, or sit next to it under
> extensions? And what should a replacement query look like, given that query
> and replacement_query are strings while a Layer 3 query is an object?
>
> I also ran 80 agent sessions: 20 analyst questions, two models, the new
> input and the previous one. Agents made no errors in the shape.
>
> Yesterday I wrote that the codes matter because the agent repairs its query
> from the refusal. The run does not show that: agents repaired from the
> message and the did-you-mean suggestion, and did about as well without
> codes. Codes may matter more to programs around the agent, and for a
> refusal to mean the same on every engine. Refusals such as E3013 and
> E_NO_PATH never came up, because the test model has one central dataset.
>
> Warm regards,
> Alex
>
> [1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
> <
> https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E
> >,
> my message of 10-07
> [2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
> <
> https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E
> >,
> the full report
>
> On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <a....@gmail.com>
> wrote:
>
> Hi all,
>
> On JB's question about engines: I'd like to offer ossie-clickhouse [1] as a
> test bed for a shared shape. I'm starting now and will report to this
> thread as results come in.
> [...]
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi all,

Following up on my message of 10-07 [1]. The ossie-clickhouse query tool
now takes the aggregation query from #246 §5.1.1 as a JSON object. Refusals
carry #246 codes in the #529 error envelope. Implementing it raised a few
questions about both drafts. The full report, with line references, is at
[2]. Will, I'm copying you because most of the questions are about #246.

The three things I promised to report:
- The executor raises 10 of the 28 #246 codes as specified, and
E_PRIMARY_KEY_REQUIRED by analogy with §4.2.
- Refusals with no #246 code get a #529 common code: UNSUPPORTED_QUERY when
the query is valid but the executor cannot compute it, QUERY_INVALID for
parse errors.
- Three kinds of queries that #246 expects answered get UNSUPPORTED_QUERY:
multi-step relationship paths, measures over different datasets in one
query, and metrics over a joined dataset's rows.

Questions for #246:
1. The object form of Order By has no key for NULLS FIRST or LAST, although
§5.1 lets an entry specify it. I added a nulls key as an extension.
2. Does E3013 apply within one measure, as the §6.2 table says, or within
one query, as Appendix A says?
3. §6.8.2 requires E3013 for two unrelated facts when Dimensions is empty,
since a Cartesian product would be wrong. But §6.6 row 4 joins one-row
totals with a CROSS JOIN, which gives one row. Which rule applies?
4. Can a path between two dimension datasets run through a fact? §6.9
builds paths from N:1 and 1:1 edges only.
5. Which code should an engine use for a query that #246 answers but the
engine does not? Semantic 5 lets engines differ on M:N edge cases, and
E3011 opts out of M:N as a whole. The executor also refuses, for example,
two measures over different datasets in one query, and finds no #246 code
for that.

Two questions for #529, when the input is a Layer 3 object. Should the #246
code replace the #529 code in error.code, or sit next to it under
extensions? And what should a replacement query look like, given that query
and replacement_query are strings while a Layer 3 query is an object?

I also ran 80 agent sessions: 20 analyst questions, two models, the new
input and the previous one. Agents made no errors in the shape.

Yesterday I wrote that the codes matter because the agent repairs its query
from the refusal. The run does not show that: agents repaired from the
message and the did-you-mean suggestion, and did about as well without
codes. Codes may matter more to programs around the agent, and for a
refusal to mean the same on every engine. Refusals such as E3013 and
E_NO_PATH never came up, because the test model has one central dataset.

Warm regards,
Alex

[1] https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm
<https://www.google.com/url?q=https://lists.apache.org/thread/ofxmoqdyjwvvj5xkc9hwvf0qt60mlhgm&source=gmail&ust=1791508588003000&sa=E>,
my message of 10-07
[2] https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1
<https://www.google.com/url?q=https://gist.github.com/ekut/7f378df5dbfae7bedc7a6ebefe4a93b1&source=gmail&ust=1791508588003000&sa=E>,
the full report

On Thu, Oct 08, 2026 04:17 AM, Aleksandr Kosachev <a....@gmail.com>
wrote:

Hi all,

On JB's question about engines: I'd like to offer ossie-clickhouse [1] as a
test bed for a shared shape. I'm starting now and will report to this
thread as results come in.
[...]

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi all,

On JB's question about engines: I'd like to offer ossie-clickhouse [1] as a
test bed for a shared shape. I'm starting now and will report to this
thread as results come in.

I plan to try two things:
1. The aggregation query from #246 §5.1.1 [2] (dimensions, measures, where,
having, order_by, limit), taken as a JSON object in query.
2. The #246 error codes, such as E_NAME_NOT_FOUND, E_NO_PATH and
E_AMBIGUOUS_PATH, returned in the #529 error envelope (error.code,
retryable, suggestions).

The report will list which #246 codes the executor raises, which of its
refusals have no code, and which queries #246 expects answered that it
still refuses.

I picked #246 because #452 points to it for Layer 3. Chris is right that
the query interface has not landed, but the shape JB describes already has
a draft there. A tool that takes that object and names the Layer 3 version
it follows does not have to define query semantics of its own.

The codes matter because the agent repairs its query from the refusal. In
Kim's experiment [3], a dataset with no relationship path was refused
instead of cross-joined. If the refusal carries a #246 code in #529's
error.code, an agent can repair the query the same way on any engine.

If #529 is narrowed, most of it carries over: the binding by
data_source_id, the result format and the diagnostics. Only the input
changes, from a native string to a Layer 3 query object.

Kind regards,
Alex Kosachev

[1] https://github.com/Dactopus/ossie-clickhouse
<https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791490479955000&sa=E>,
an Ossie executor on ClickHouse
[2] https://github.com/apache/ossie/pull/246
<https://www.google.com/url?q=https://github.com/apache/ossie/pull/246&source=gmail&ust=1791490479955000&sa=E>,
Foundational Semantics spec (Layer 3)
[3] https://lists.apache.org/thread/wnfkdwbx7qjkh8mkp0zmn875gfbodyxw
<https://www.google.com/url?q=https://lists.apache.org/thread/wnfkdwbx7qjkh8mkp0zmn875gfbodyxw&source=gmail&ust=1791490479955000&sa=E>,
Kim's Layer 3 experiment

On Wed, Oct 07, 2026 07:59 PM, Jean-Baptiste Onofré <jb...@nanthrax.net> wrote:

Hi all,

I think Chris's question about how tightly execute_query ties to Ossie
semantics decides where this should live.

Ossie is an interchange spec for semantic models (at least today ;)). A
running MCP server brings in execution, connections, auth and a fast-moving
protocol surface. I would rather not take on in core, so my preference is
for MCP to live in engines and catalogs, as I said earlier. The spec could
make sense (like the REST API) but not the server implementation.

The semantic question matters more than the location, though. I see two
different things here:
1. The declarative queries (metrics, dimensions, filters) are resolved
against the model. That's the way Aleksandr's ossie-clickhouse is working.
The semantics come from the model, so two implementations should return the
same answer for the same metric. That is a contract Ossie can legitimately
define.
2. This is a generic agent to query engine interface. Ossie could not say
whether two servers compute a metric the same way, so the spec would add
nothing here.

If we want anything shared, I would keep it small and non-normative. It
would be a declarative query shape plus a tool schema for model discovery,
published as a reference or in a separate repo, with no server code in this
repository. Discovery tools can stay optional.

Markus, would you be open to narrow PR #529 to the declarative case? I
would also like to hear from other engine and catalog people on whether
they would implement a shared shape.

Regards
JB

On Wed, Oct 7, 2026 at 11:02 AM Markus Cozowicz <m...@cozowicz.at> wrote:

> Hi Chris,
>
> Maybe being a bit pragmatic here, but it a) might take a while until all
> vendors have conversions and b) it's most likely that each vendor might
> have some unique features that are not part of the standard yet, thus
> allowing all registered Ossie dialects seems like a good starting point to
> me.
>
> From a goal perspective, if execute_query (or whatever we call it), makes
> it into the Ossie standard and subsequently into the training corpus for
> LLMs, we all don't have to teach the various agents how to do query
> execution, as it would just be common knowledge (and hopefully comes with
> all the benefits: less tokens, lower latency, higher accuracy,...)
>
> Markus
> ________________________________
> From: Chris Eubank via dev <de...@ossie.apache.org>
> Sent: Wednesday, October 7, 2026 1:36 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi Markus, Alex, JB,
>
> Thanks for opening #529 - it’s great to have something concrete to
discuss.
>
> I'm curious about the intended direction of execute_query. To what extent
> is the contract meant to be tied to Ossie semantics? Is the goal to guide
> clients toward portable Ossie query interfaces, or also to standardize the
> execution of vendor-specific query languages compiled outside Ossie? #529
> currently seems closer to the latter. Its ties to Ossie also look fairly
> loose - the model identity is optional, and the result schema/logical
types
> appear to be the main semantic link. Is that intentionally the level of
> coupling to Ossie?
>
> I’m asking because these seem to give pretty different portability
> guarantees. If the caller sends an engine-native query, do two tools
> consuming the same Ossie model each need to translate the semantic
> definitions themselves? If so, where do we ensure they preserve the same
> semantics?
>
> The implementations we have today seem to illustrate the distinction.
> Alex’s ossie-clickhouse takes a query in Ossie’s in-flight Layer 3 terms
> (metrics, dimensions, filters) and compiles it, whereas #529 accepts a
> native query string. On binding, they seem aligned: the model carries no
> connection, the server owns it, and the agent sees the exposed tools. How
> do you see these two approaches fitting under a common contract?
>
> This also makes me wonder about scope. A vendor-agnostic execute_query
> interface for agents seems useful, but if the query itself is
> engine-native, I’m not yet sure what makes the contract specifically Ossie
> rather than a more general agent-to-query-engine interface.
>
> JB, this may connect to your point about MCP looking closer to an
> engine/catalog concern - curious how you see the boundary.
>
> Kind regards,
> Chris
>
> On Tue, Oct 6, 2026 at 12:24 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > Hi Alex,
> >
> > Good see we're following similar patterns.
> >
> > Input of execute_query: I'm not sure we should limit the input of
execute
> > query to a specific layer or dialect. Some of our implementations
support
> > 4-5 different query languages, depending on the targeted data engine.
> >
> > As for filter values: the canonical example: SELECT * FROM X WHERE
> > CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as
> > the filter value. In a simplified version, you can check if the results
> are
> > empty and the using backend logic figure what are alternative values you
> > present to the agent.
> >
> > The below seems very similar to how we do things.
> > - The model carries no connection.
> > - The server starts with a model and a database connection.
> > - The agent sees only the tools.
> >
> > I created a PR https://github.com/apache/ossie/pull/529
<https://www.google.com/url?q=https://github.com/apache/ossie/pull/529&source=gmail&ust=1791490479955000&sa=E>
to get us
> started
> > on more details.
> >
> >
> > Markus
> > ________________________________
> > From: Aleksandr Kosachev <a....@gmail.com>
> > Sent: Tuesday, October 6, 2026 1:22 AM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> > Hi Markus, JB,
> >
> > One gap the thread has not named yet. The REST API spec proposed on
09-23
> > covers two things: managing models and binding a model to physical
> > entities. It does not cover running a query and returning data. That is
> > where execute_query starts. So an MCP spec next to the REST spec would
> not
> > duplicate it. It would fill the missing step.
> >
> > Markus asked how an agent that found a model knows how to run a query
> > against the source.
> > In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain,
> it
> > does not need to know:
> > - The model carries no connection.
> > - The server starts with a model and a database connection.
> > - The agent sees only the tools.
> >
> > This reconciles the two positions in the thread. Ossie specifies the
tool
> > contract. The engine or catalog that serves it owns the binding.
> >
> > One question needs an answer before a spec: what is the input of
> > execute_query? PR #452 names two query interfaces:
> > - Layer 2: SQL with MEASURE().
> > - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
> >
> > execute_query could accept either. ossie-clickhouse takes the Layer 3
> form.
> > When its planner refuses a query, the refusal names the cause.
> > For example, "unknown metric 'x' (did you mean: ...)", or which metric
> > would be counted once per joined row.
> > The agent reads that and repairs the query.
> > Markus listed four outputs: data, errors, suggestions, filter value
> > alternatives.
> > ossie-clickhouse returns data, errors and name suggestions, plus the
> > generated SQL. It has no filter value alternatives yet.
> > I would like to hear how those look on your side.
> >
> > On query_model and the harness dependency, ossie-clickhouse confirms it.
> > Its overview tool returns datasets and metrics without fields.
> > Agents then call a search tool and a describe-one-object tool before
> every
> > query. A harness with the file on disk, as Markus described, does not
> need
> > them. So discovery tools belong in the spec as optional, not dropped.
> >
> > Warm regards,
> > Alex Kosachev
> >
> > [1] https://github.com/Dactopus/ossie-clickhouse
<https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791490479956000&sa=E>
> > <
> >
>
https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
<https://www.google.com/url?q=https://www.google.com/url?q%3Dhttps://github.com/Dactopus/ossie-clickhouse%26source%3Dgmail%26ust%3D1791327988788000%26sa%3DE&source=gmail&ust=1791490479956000&sa=E>
> > >
> >
> > On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
> >
> > > I'm curious about the expectations here. If e.g. GitHub copilot
> discovers
> > > an Ossie semantic model, how would it know how to connect/execute a
> query
> > > against the data source (e.g. PBI semantic model, Databricks,
> Snowflake,
> > > ...)? Is the expectation that the harness would get additional
> > instructions
> > > anyway?
> > >
> > > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> > > implementation.
> > > ________________________________
> > > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> > > Sent: Monday, October 5, 2026 8:25 PM
> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
should
> > > live) [ossie]
> > >
> > > Hi
> > >
> > > It's not clear to me if we are discussing the specification or the
> > > implementation. MCP implementation in Ossie: I'm not convinced; it
> seems
> > > more like an engine/catalog implementation.
> > > The spec could make sense.
> > >
> > > But do we really need the MCP spec in Ossie for agent model discovery?
> > I'm
> > > not sure. That's actually one of the reasons for the REST API Spec too
> > > (model discovery, versioning, etc).
> > > We could have the MCP spec alongside the REST API Spec.
> > >
> > > Also, MCP is good to have but clearly not a requirement for the agent
> > (API
> > > or CLI are classic data providers for agent, including operations).
> > >
> > > Regards
> > > JB
> > >
> > > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> > >
> > > > I'd like to provide a slightly different pitch for an MCP server
spec
> > > (not
> > > > necessarily the implementation). From my perspective, Ossies biggest
> > > value
> > > > prop is related to empowering agents to understand semantics and
> > > > subsequently build queries to retrieve data.
> > > > Most important is probably an execute_query MCP tool.
> > > >
> > > >
> > > > *
> > > > Input: data source id, query
> > > > *
> > > > Output
> > > > *
> > > > Data
> > > > *
> > > > Errors
> > > > *
> > > > Suggestions
> > > > *
> > > > Filter value alternatives
> > > > *
> > > > ...
> > > >
> > > > This operates under the assumption that the agent is able to
> > discover/see
> > > > the Ossie semantic model (e.g. using
> > > http://agenticresourcediscovery.org/
<https://www.google.com/url?q=http://agenticresourcediscovery.org/&source=gmail&ust=1791490479956000&sa=E>
).
> > > > We already have a few variants of execute_query tool across our
> > engines.
> > > >
> > > > I'm not sure about query_model (yet), at least for now we see a
> > > dependency
> > > > on the agent harness. Some harness can only digest a small view of
> the
> > > > semantic model and thus need tools to progressively discover, others
> > > (like
> > > > e.g. GitHub copilot cli), don't need any of that, as they'll have
> full
> > > > filesystem access, download/query local.
> > > >
> > > > ________________________________
> > > > From: jbonofre (via GitHub) <gi...@apache.org>
> > > > Sent: Monday, October 5, 2026 8:01 AM
> > > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
> should
> > > > live) [ossie]
> > > >
> > > >
> > > > GitHub user jbonofre added a comment to the discussion: Proposal: an
> > MCP
> > > > server for Ossie (and where it should live)
> > > >
> > > > I'm not sure we need MCP server at all in Ossie. That could be
> > something
> > > > catalogs can do.
> > > >
> > > > I would rather not consider MCP service directly in Ossie for now (I
> > > don't
> > > > see the value).
> > > >
> > > > GitHub link:
> > > >
> > >
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
<https://www.google.com/url?q=https://github.com/apache/ossie/discussions/510%23discussioncomment-18754069&source=gmail&ust=1791490479956000&sa=E>
> > > >
> > > > ----
> > > > This is an automatically sent email for dev@ossie.apache.org.
> > > > To unsubscribe, please send an email to:
> > > dev-unsubscribe@ossie.apache.org
> > > >
> > > >
> > >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Jean-Baptiste Onofré <jb...@nanthrax.net>.
Hi all,

I think Chris's question about how tightly execute_query ties to Ossie
semantics decides where this should live.

Ossie is an interchange spec for semantic models (at least today ;)). A
running MCP server brings in execution, connections, auth and a fast-moving
protocol surface. I would rather not take on in core, so my preference is
for MCP to live in engines and catalogs, as I said earlier. The spec could
make sense (like the REST API) but not the server implementation.

The semantic question matters more than the location, though. I see two
different things here:
1. The declarative queries (metrics, dimensions, filters) are resolved
against the model. That's the way Aleksandr's ossie-clickhouse is working.
The semantics come from the model, so two implementations should return the
same answer for the same metric. That is a contract Ossie can legitimately
define.
2. This is a generic agent to query engine interface. Ossie could not say
whether two servers compute a metric the same way, so the spec would add
nothing here.

If we want anything shared, I would keep it small and non-normative. It
would be a declarative query shape plus a tool schema for model discovery,
published as a reference or in a separate repo, with no server code in this
repository. Discovery tools can stay optional.

Markus, would you be open to narrow PR #529 to the declarative case? I
would also like to hear from other engine and catalog people on whether
they would implement a shared shape.

Regards
JB

On Wed, Oct 7, 2026 at 11:02 AM Markus Cozowicz <m...@cozowicz.at> wrote:

> Hi Chris,
>
> Maybe being a bit pragmatic here, but it a) might take a while until all
> vendors have conversions and b) it's most likely that each vendor might
> have some unique features that are not part of the standard yet, thus
> allowing all registered Ossie dialects seems like a good starting point to
> me.
>
> From a goal perspective, if execute_query (or whatever we call it), makes
> it into the Ossie standard and subsequently into the training corpus for
> LLMs, we all don't have to teach the various agents how to do query
> execution, as it would just be common knowledge (and hopefully comes with
> all the benefits: less tokens, lower latency, higher accuracy,...)
>
> Markus
> ________________________________
> From: Chris Eubank via dev <de...@ossie.apache.org>
> Sent: Wednesday, October 7, 2026 1:36 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi Markus, Alex, JB,
>
> Thanks for opening #529 - it’s great to have something concrete to discuss.
>
> I'm curious about the intended direction of execute_query. To what extent
> is the contract meant to be tied to Ossie semantics? Is the goal to guide
> clients toward portable Ossie query interfaces, or also to standardize the
> execution of vendor-specific query languages compiled outside Ossie? #529
> currently seems closer to the latter.  Its ties to Ossie also look fairly
> loose - the model identity is optional, and the result schema/logical types
> appear to be the main semantic link. Is that intentionally the level of
> coupling to Ossie?
>
> I’m asking because these seem to give pretty different portability
> guarantees. If the caller sends an engine-native query, do two tools
> consuming the same Ossie model each need to translate the semantic
> definitions themselves? If so, where do we ensure they preserve the same
> semantics?
>
> The implementations we have today seem to illustrate the distinction.
> Alex’s ossie-clickhouse takes a query in Ossie’s in-flight Layer 3 terms
> (metrics, dimensions, filters) and compiles it, whereas #529 accepts a
> native query string. On binding, they seem aligned: the model carries no
> connection, the server owns it, and the agent sees the exposed tools. How
> do you see these two approaches fitting under a common contract?
>
> This also makes me wonder about scope. A vendor-agnostic execute_query
> interface for agents seems useful, but if the query itself is
> engine-native, I’m not yet sure what makes the contract specifically Ossie
> rather than a more general agent-to-query-engine interface.
>
> JB, this may connect to your point about MCP looking closer to an
> engine/catalog concern - curious how you see the boundary.
>
> Kind regards,
> Chris
>
> On Tue, Oct 6, 2026 at 12:24 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > Hi Alex,
> >
> > Good see we're following similar patterns.
> >
> > Input of execute_query: I'm not sure we should limit the input of execute
> > query to a specific layer or dialect. Some of our implementations support
> > 4-5 different query languages, depending on the targeted data engine.
> >
> > As for filter values: the canonical example: SELECT * FROM X WHERE
> > CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as
> > the filter value. In a simplified version, you can check if the results
> are
> > empty and the using backend logic figure what are alternative values you
> > present to the agent.
> >
> > The below seems very similar to how we do things.
> > - The model carries no connection.
> > - The server starts with a model and a database connection.
> > - The agent sees only the tools.
> >
> > I created a PR https://github.com/apache/ossie/pull/529 to get us
> started
> > on more details.
> >
> >
> > Markus
> > ________________________________
> > From: Aleksandr Kosachev <a....@gmail.com>
> > Sent: Tuesday, October 6, 2026 1:22 AM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> > Hi Markus, JB,
> >
> > One gap the thread has not named yet. The REST API spec proposed on 09-23
> > covers two things: managing models and binding a model to physical
> > entities. It does not cover running a query and returning data. That is
> > where execute_query starts. So an MCP spec next to the REST spec would
> not
> > duplicate it. It would fill the missing step.
> >
> > Markus asked how an agent that found a model knows how to run a query
> > against the source.
> > In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain,
> it
> > does not need to know:
> > - The model carries no connection.
> > - The server starts with a model and a database connection.
> > - The agent sees only the tools.
> >
> > This reconciles the two positions in the thread. Ossie specifies the tool
> > contract. The engine or catalog that serves it owns the binding.
> >
> > One question needs an answer before a spec: what is the input of
> > execute_query? PR #452 names two query interfaces:
> > - Layer 2: SQL with MEASURE().
> > - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
> >
> > execute_query could accept either. ossie-clickhouse takes the Layer 3
> form.
> > When its planner refuses a query, the refusal names the cause.
> > For example, "unknown metric 'x' (did you mean: ...)", or which metric
> > would be counted once per joined row.
> > The agent reads that and repairs the query.
> > Markus listed four outputs: data, errors, suggestions, filter value
> > alternatives.
> > ossie-clickhouse returns data, errors and name suggestions, plus the
> > generated SQL. It has no filter value alternatives yet.
> > I would like to hear how those look on your side.
> >
> > On query_model and the harness dependency, ossie-clickhouse confirms it.
> > Its overview tool returns datasets and metrics without fields.
> > Agents then call a search tool and a describe-one-object tool before
> every
> > query. A harness with the file on disk, as Markus described, does not
> need
> > them. So discovery tools belong in the spec as optional, not dropped.
> >
> > Warm regards,
> > Alex Kosachev
> >
> > [1] https://github.com/Dactopus/ossie-clickhouse
> > <
> >
> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
> > >
> >
> > On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
> >
> > > I'm curious about the expectations here. If e.g. GitHub copilot
> discovers
> > > an Ossie semantic model, how would it know how to connect/execute a
> query
> > > against the data source (e.g. PBI semantic model, Databricks,
> Snowflake,
> > > ...)? Is the expectation that the harness would get additional
> > instructions
> > > anyway?
> > >
> > > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> > > implementation.
> > > ________________________________
> > > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> > > Sent: Monday, October 5, 2026 8:25 PM
> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > > live) [ossie]
> > >
> > > Hi
> > >
> > > It's not clear to me if we are discussing the specification or the
> > > implementation. MCP implementation in Ossie: I'm not convinced; it
> seems
> > > more like an engine/catalog implementation.
> > > The spec could make sense.
> > >
> > > But do we really need the MCP spec in Ossie for agent model discovery?
> > I'm
> > > not sure. That's actually one of the reasons for the REST API Spec too
> > > (model discovery, versioning, etc).
> > > We could have the MCP spec alongside the REST API Spec.
> > >
> > > Also, MCP is good to have but clearly not a requirement for the agent
> > (API
> > > or CLI are classic data providers for agent, including operations).
> > >
> > > Regards
> > > JB
> > >
> > > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> > >
> > > > I'd like to provide a slightly different pitch for an MCP server spec
> > > (not
> > > > necessarily the implementation). From my perspective, Ossies biggest
> > > value
> > > > prop is related to empowering agents to understand semantics and
> > > > subsequently build queries to retrieve data.
> > > > Most important is probably an execute_query MCP tool.
> > > >
> > > >
> > > >   *
> > > > Input: data source id, query
> > > >   *
> > > > Output
> > > >      *
> > > > Data
> > > >      *
> > > > Errors
> > > >      *
> > > > Suggestions
> > > >      *
> > > > Filter value alternatives
> > > >      *
> > > > ...
> > > >
> > > > This operates under the assumption that the agent is able to
> > discover/see
> > > > the Ossie semantic model (e.g. using
> > > http://agenticresourcediscovery.org/).
> > > > We already have a few variants of execute_query tool across our
> > engines.
> > > >
> > > > I'm not sure about query_model (yet), at least for now we see a
> > > dependency
> > > > on the agent harness. Some harness can only digest a small view of
> the
> > > > semantic model and thus need tools to progressively discover, others
> > > (like
> > > > e.g. GitHub copilot cli), don't need any of that, as they'll have
> full
> > > > filesystem access, download/query local.
> > > >
> > > > ________________________________
> > > > From: jbonofre (via GitHub) <gi...@apache.org>
> > > > Sent: Monday, October 5, 2026 8:01 AM
> > > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it
> should
> > > > live) [ossie]
> > > >
> > > >
> > > > GitHub user jbonofre added a comment to the discussion: Proposal: an
> > MCP
> > > > server for Ossie (and where it should live)
> > > >
> > > > I'm not sure we need MCP server at all in Ossie. That could be
> > something
> > > > catalogs can do.
> > > >
> > > > I would rather not consider MCP service directly in Ossie for now (I
> > > don't
> > > > see the value).
> > > >
> > > > GitHub link:
> > > >
> > >
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> > > >
> > > > ----
> > > > This is an automatically sent email for dev@ossie.apache.org.
> > > > To unsubscribe, please send an email to:
> > > dev-unsubscribe@ossie.apache.org
> > > >
> > > >
> > >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
Hi Chris,

Maybe being a bit pragmatic here, but it a) might take a while until all vendors have conversions and b) it's most likely that each vendor might have some unique features that are not part of the standard yet, thus allowing all registered Ossie dialects seems like a good starting point to me.

From a goal perspective, if execute_query (or whatever we call it), makes it into the Ossie standard and subsequently into the training corpus for LLMs, we all don't have to teach the various agents how to do query execution, as it would just be common knowledge (and hopefully comes with all the benefits: less tokens, lower latency, higher accuracy,...)

Markus
________________________________
From: Chris Eubank via dev <de...@ossie.apache.org>
Sent: Wednesday, October 7, 2026 1:36 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Hi Markus, Alex, JB,

Thanks for opening #529 - it’s great to have something concrete to discuss.

I'm curious about the intended direction of execute_query. To what extent
is the contract meant to be tied to Ossie semantics? Is the goal to guide
clients toward portable Ossie query interfaces, or also to standardize the
execution of vendor-specific query languages compiled outside Ossie? #529
currently seems closer to the latter.  Its ties to Ossie also look fairly
loose - the model identity is optional, and the result schema/logical types
appear to be the main semantic link. Is that intentionally the level of
coupling to Ossie?

I’m asking because these seem to give pretty different portability
guarantees. If the caller sends an engine-native query, do two tools
consuming the same Ossie model each need to translate the semantic
definitions themselves? If so, where do we ensure they preserve the same
semantics?

The implementations we have today seem to illustrate the distinction.
Alex’s ossie-clickhouse takes a query in Ossie’s in-flight Layer 3 terms
(metrics, dimensions, filters) and compiles it, whereas #529 accepts a
native query string. On binding, they seem aligned: the model carries no
connection, the server owns it, and the agent sees the exposed tools. How
do you see these two approaches fitting under a common contract?

This also makes me wonder about scope. A vendor-agnostic execute_query
interface for agents seems useful, but if the query itself is
engine-native, I’m not yet sure what makes the contract specifically Ossie
rather than a more general agent-to-query-engine interface.

JB, this may connect to your point about MCP looking closer to an
engine/catalog concern - curious how you see the boundary.

Kind regards,
Chris

On Tue, Oct 6, 2026 at 12:24 PM Markus Cozowicz <m...@cozowicz.at> wrote:

> Hi Alex,
>
> Good see we're following similar patterns.
>
> Input of execute_query: I'm not sure we should limit the input of execute
> query to a specific layer or dialect. Some of our implementations support
> 4-5 different query languages, depending on the targeted data engine.
>
> As for filter values: the canonical example: SELECT * FROM X WHERE
> CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as
> the filter value. In a simplified version, you can check if the results are
> empty and the using backend logic figure what are alternative values you
> present to the agent.
>
> The below seems very similar to how we do things.
> - The model carries no connection.
> - The server starts with a model and a database connection.
> - The agent sees only the tools.
>
> I created a PR https://github.com/apache/ossie/pull/529 to get us started
> on more details.
>
>
> Markus
> ________________________________
> From: Aleksandr Kosachev <a....@gmail.com>
> Sent: Tuesday, October 6, 2026 1:22 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi Markus, JB,
>
> One gap the thread has not named yet. The REST API spec proposed on 09-23
> covers two things: managing models and binding a model to physical
> entities. It does not cover running a query and returning data. That is
> where execute_query starts. So an MCP spec next to the REST spec would not
> duplicate it. It would fill the missing step.
>
> Markus asked how an agent that found a model knows how to run a query
> against the source.
> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
> does not need to know:
> - The model carries no connection.
> - The server starts with a model and a database connection.
> - The agent sees only the tools.
>
> This reconciles the two positions in the thread. Ossie specifies the tool
> contract. The engine or catalog that serves it owns the binding.
>
> One question needs an answer before a spec: what is the input of
> execute_query? PR #452 names two query interfaces:
> - Layer 2: SQL with MEASURE().
> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>
> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
> When its planner refuses a query, the refusal names the cause.
> For example, "unknown metric 'x' (did you mean: ...)", or which metric
> would be counted once per joined row.
> The agent reads that and repairs the query.
> Markus listed four outputs: data, errors, suggestions, filter value
> alternatives.
> ossie-clickhouse returns data, errors and name suggestions, plus the
> generated SQL. It has no filter value alternatives yet.
> I would like to hear how those look on your side.
>
> On query_model and the harness dependency, ossie-clickhouse confirms it.
> Its overview tool returns datasets and metrics without fields.
> Agents then call a search tool and a describe-one-object tool before every
> query. A harness with the file on disk, as Markus described, does not need
> them. So discovery tools belong in the spec as optional, not dropped.
>
> Warm regards,
> Alex Kosachev
>
> [1] https://github.com/Dactopus/ossie-clickhouse
> <
> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
> >
>
> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > I'm curious about the expectations here. If e.g. GitHub copilot discovers
> > an Ossie semantic model, how would it know how to connect/execute a query
> > against the data source (e.g. PBI semantic model, Databricks, Snowflake,
> > ...)? Is the expectation that the harness would get additional
> instructions
> > anyway?
> >
> > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> > implementation.
> > ________________________________
> > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> > Sent: Monday, October 5, 2026 8:25 PM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> > Hi
> >
> > It's not clear to me if we are discussing the specification or the
> > implementation. MCP implementation in Ossie: I'm not convinced; it seems
> > more like an engine/catalog implementation.
> > The spec could make sense.
> >
> > But do we really need the MCP spec in Ossie for agent model discovery?
> I'm
> > not sure. That's actually one of the reasons for the REST API Spec too
> > (model discovery, versioning, etc).
> > We could have the MCP spec alongside the REST API Spec.
> >
> > Also, MCP is good to have but clearly not a requirement for the agent
> (API
> > or CLI are classic data providers for agent, including operations).
> >
> > Regards
> > JB
> >
> > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> >
> > > I'd like to provide a slightly different pitch for an MCP server spec
> > (not
> > > necessarily the implementation). From my perspective, Ossies biggest
> > value
> > > prop is related to empowering agents to understand semantics and
> > > subsequently build queries to retrieve data.
> > > Most important is probably an execute_query MCP tool.
> > >
> > >
> > >   *
> > > Input: data source id, query
> > >   *
> > > Output
> > >      *
> > > Data
> > >      *
> > > Errors
> > >      *
> > > Suggestions
> > >      *
> > > Filter value alternatives
> > >      *
> > > ...
> > >
> > > This operates under the assumption that the agent is able to
> discover/see
> > > the Ossie semantic model (e.g. using
> > http://agenticresourcediscovery.org/).
> > > We already have a few variants of execute_query tool across our
> engines.
> > >
> > > I'm not sure about query_model (yet), at least for now we see a
> > dependency
> > > on the agent harness. Some harness can only digest a small view of the
> > > semantic model and thus need tools to progressively discover, others
> > (like
> > > e.g. GitHub copilot cli), don't need any of that, as they'll have full
> > > filesystem access, download/query local.
> > >
> > > ________________________________
> > > From: jbonofre (via GitHub) <gi...@apache.org>
> > > Sent: Monday, October 5, 2026 8:01 AM
> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > > live) [ossie]
> > >
> > >
> > > GitHub user jbonofre added a comment to the discussion: Proposal: an
> MCP
> > > server for Ossie (and where it should live)
> > >
> > > I'm not sure we need MCP server at all in Ossie. That could be
> something
> > > catalogs can do.
> > >
> > > I would rather not consider MCP service directly in Ossie for now (I
> > don't
> > > see the value).
> > >
> > > GitHub link:
> > >
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> > >
> > > ----
> > > This is an automatically sent email for dev@ossie.apache.org.
> > > To unsubscribe, please send an email to:
> > dev-unsubscribe@ossie.apache.org
> > >
> > >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Chris Eubank via dev <de...@ossie.apache.org>.
Hi Markus, Alex, JB,

Thanks for opening #529 - it’s great to have something concrete to discuss.

I'm curious about the intended direction of execute_query. To what extent
is the contract meant to be tied to Ossie semantics? Is the goal to guide
clients toward portable Ossie query interfaces, or also to standardize the
execution of vendor-specific query languages compiled outside Ossie? #529
currently seems closer to the latter.  Its ties to Ossie also look fairly
loose - the model identity is optional, and the result schema/logical types
appear to be the main semantic link. Is that intentionally the level of
coupling to Ossie?

I’m asking because these seem to give pretty different portability
guarantees. If the caller sends an engine-native query, do two tools
consuming the same Ossie model each need to translate the semantic
definitions themselves? If so, where do we ensure they preserve the same
semantics?

The implementations we have today seem to illustrate the distinction.
Alex’s ossie-clickhouse takes a query in Ossie’s in-flight Layer 3 terms
(metrics, dimensions, filters) and compiles it, whereas #529 accepts a
native query string. On binding, they seem aligned: the model carries no
connection, the server owns it, and the agent sees the exposed tools. How
do you see these two approaches fitting under a common contract?

This also makes me wonder about scope. A vendor-agnostic execute_query
interface for agents seems useful, but if the query itself is
engine-native, I’m not yet sure what makes the contract specifically Ossie
rather than a more general agent-to-query-engine interface.

JB, this may connect to your point about MCP looking closer to an
engine/catalog concern - curious how you see the boundary.

Kind regards,
Chris

On Tue, Oct 6, 2026 at 12:24 PM Markus Cozowicz <m...@cozowicz.at> wrote:

> Hi Alex,
>
> Good see we're following similar patterns.
>
> Input of execute_query: I'm not sure we should limit the input of execute
> query to a specific layer or dialect. Some of our implementations support
> 4-5 different query languages, depending on the targeted data engine.
>
> As for filter values: the canonical example: SELECT * FROM X WHERE
> CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as
> the filter value. In a simplified version, you can check if the results are
> empty and the using backend logic figure what are alternative values you
> present to the agent.
>
> The below seems very similar to how we do things.
> - The model carries no connection.
> - The server starts with a model and a database connection.
> - The agent sees only the tools.
>
> I created a PR https://github.com/apache/ossie/pull/529 to get us started
> on more details.
>
>
> Markus
> ________________________________
> From: Aleksandr Kosachev <a....@gmail.com>
> Sent: Tuesday, October 6, 2026 1:22 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi Markus, JB,
>
> One gap the thread has not named yet. The REST API spec proposed on 09-23
> covers two things: managing models and binding a model to physical
> entities. It does not cover running a query and returning data. That is
> where execute_query starts. So an MCP spec next to the REST spec would not
> duplicate it. It would fill the missing step.
>
> Markus asked how an agent that found a model knows how to run a query
> against the source.
> In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
> does not need to know:
> - The model carries no connection.
> - The server starts with a model and a database connection.
> - The agent sees only the tools.
>
> This reconciles the two positions in the thread. Ossie specifies the tool
> contract. The engine or catalog that serves it owns the binding.
>
> One question needs an answer before a spec: what is the input of
> execute_query? PR #452 names two query interfaces:
> - Layer 2: SQL with MEASURE().
> - Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.
>
> execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
> When its planner refuses a query, the refusal names the cause.
> For example, "unknown metric 'x' (did you mean: ...)", or which metric
> would be counted once per joined row.
> The agent reads that and repairs the query.
> Markus listed four outputs: data, errors, suggestions, filter value
> alternatives.
> ossie-clickhouse returns data, errors and name suggestions, plus the
> generated SQL. It has no filter value alternatives yet.
> I would like to hear how those look on your side.
>
> On query_model and the harness dependency, ossie-clickhouse confirms it.
> Its overview tool returns datasets and metrics without fields.
> Agents then call a search tool and a describe-one-object tool before every
> query. A harness with the file on disk, as Markus described, does not need
> them. So discovery tools belong in the spec as optional, not dropped.
>
> Warm regards,
> Alex Kosachev
>
> [1] https://github.com/Dactopus/ossie-clickhouse
> <
> https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E
> >
>
> On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > I'm curious about the expectations here. If e.g. GitHub copilot discovers
> > an Ossie semantic model, how would it know how to connect/execute a query
> > against the data source (e.g. PBI semantic model, Databricks, Snowflake,
> > ...)? Is the expectation that the harness would get additional
> instructions
> > anyway?
> >
> > Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> > implementation.
> > ________________________________
> > From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> > Sent: Monday, October 5, 2026 8:25 PM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> > Hi
> >
> > It's not clear to me if we are discussing the specification or the
> > implementation. MCP implementation in Ossie: I'm not convinced; it seems
> > more like an engine/catalog implementation.
> > The spec could make sense.
> >
> > But do we really need the MCP spec in Ossie for agent model discovery?
> I'm
> > not sure. That's actually one of the reasons for the REST API Spec too
> > (model discovery, versioning, etc).
> > We could have the MCP spec alongside the REST API Spec.
> >
> > Also, MCP is good to have but clearly not a requirement for the agent
> (API
> > or CLI are classic data providers for agent, including operations).
> >
> > Regards
> > JB
> >
> > On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
> >
> > > I'd like to provide a slightly different pitch for an MCP server spec
> > (not
> > > necessarily the implementation). From my perspective, Ossies biggest
> > value
> > > prop is related to empowering agents to understand semantics and
> > > subsequently build queries to retrieve data.
> > > Most important is probably an execute_query MCP tool.
> > >
> > >
> > >   *
> > > Input: data source id, query
> > >   *
> > > Output
> > >      *
> > > Data
> > >      *
> > > Errors
> > >      *
> > > Suggestions
> > >      *
> > > Filter value alternatives
> > >      *
> > > ...
> > >
> > > This operates under the assumption that the agent is able to
> discover/see
> > > the Ossie semantic model (e.g. using
> > http://agenticresourcediscovery.org/).
> > > We already have a few variants of execute_query tool across our
> engines.
> > >
> > > I'm not sure about query_model (yet), at least for now we see a
> > dependency
> > > on the agent harness. Some harness can only digest a small view of the
> > > semantic model and thus need tools to progressively discover, others
> > (like
> > > e.g. GitHub copilot cli), don't need any of that, as they'll have full
> > > filesystem access, download/query local.
> > >
> > > ________________________________
> > > From: jbonofre (via GitHub) <gi...@apache.org>
> > > Sent: Monday, October 5, 2026 8:01 AM
> > > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > > live) [ossie]
> > >
> > >
> > > GitHub user jbonofre added a comment to the discussion: Proposal: an
> MCP
> > > server for Ossie (and where it should live)
> > >
> > > I'm not sure we need MCP server at all in Ossie. That could be
> something
> > > catalogs can do.
> > >
> > > I would rather not consider MCP service directly in Ossie for now (I
> > don't
> > > see the value).
> > >
> > > GitHub link:
> > >
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> > >
> > > ----
> > > This is an automatically sent email for dev@ossie.apache.org.
> > > To unsubscribe, please send an email to:
> > dev-unsubscribe@ossie.apache.org
> > >
> > >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
Hi Alex,

Good see we're following similar patterns.

Input of execute_query: I'm not sure we should limit the input of execute query to a specific layer or dialect. Some of our implementations support 4-5 different query languages, depending on the targeted data engine.

As for filter values: the canonical example: SELECT * FROM X WHERE CompanyName = 'Coca & Cola' , but really you need 'Coca-Cola Company' as the filter value. In a simplified version, you can check if the results are empty and the using backend logic figure what are alternative values you present to the agent.

The below seems very similar to how we do things.
- The model carries no connection.
- The server starts with a model and a database connection.
- The agent sees only the tools.

I created a PR https://github.com/apache/ossie/pull/529 to get us started on more details.


Markus
________________________________
From: Aleksandr Kosachev <a....@gmail.com>
Sent: Tuesday, October 6, 2026 1:22 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Hi Markus, JB,

One gap the thread has not named yet. The REST API spec proposed on 09-23
covers two things: managing models and binding a model to physical
entities. It does not cover running a query and returning data. That is
where execute_query starts. So an MCP spec next to the REST spec would not
duplicate it. It would fill the missing step.

Markus asked how an agent that found a model knows how to run a query
against the source.
In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
does not need to know:
- The model carries no connection.
- The server starts with a model and a database connection.
- The agent sees only the tools.

This reconciles the two positions in the thread. Ossie specifies the tool
contract. The engine or catalog that serves it owns the binding.

One question needs an answer before a spec: what is the input of
execute_query? PR #452 names two query interfaces:
- Layer 2: SQL with MEASURE().
- Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.

execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
When its planner refuses a query, the refusal names the cause.
For example, "unknown metric 'x' (did you mean: ...)", or which metric
would be counted once per joined row.
The agent reads that and repairs the query.
Markus listed four outputs: data, errors, suggestions, filter value
alternatives.
ossie-clickhouse returns data, errors and name suggestions, plus the
generated SQL. It has no filter value alternatives yet.
I would like to hear how those look on your side.

On query_model and the harness dependency, ossie-clickhouse confirms it.
Its overview tool returns datasets and metrics without fields.
Agents then call a search tool and a describe-one-object tool before every
query. A harness with the file on disk, as Markus described, does not need
them. So discovery tools belong in the spec as optional, not dropped.

Warm regards,
Alex Kosachev

[1] https://github.com/Dactopus/ossie-clickhouse
<https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E>

On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:

> I'm curious about the expectations here. If e.g. GitHub copilot discovers
> an Ossie semantic model, how would it know how to connect/execute a query
> against the data source (e.g. PBI semantic model, Databricks, Snowflake,
> ...)? Is the expectation that the harness would get additional instructions
> anyway?
>
> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> implementation.
> ________________________________
> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> Sent: Monday, October 5, 2026 8:25 PM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi
>
> It's not clear to me if we are discussing the specification or the
> implementation. MCP implementation in Ossie: I'm not convinced; it seems
> more like an engine/catalog implementation.
> The spec could make sense.
>
> But do we really need the MCP spec in Ossie for agent model discovery? I'm
> not sure. That's actually one of the reasons for the REST API Spec too
> (model discovery, versioning, etc).
> We could have the MCP spec alongside the REST API Spec.
>
> Also, MCP is good to have but clearly not a requirement for the agent (API
> or CLI are classic data providers for agent, including operations).
>
> Regards
> JB
>
> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > I'd like to provide a slightly different pitch for an MCP server spec
> (not
> > necessarily the implementation). From my perspective, Ossies biggest
> value
> > prop is related to empowering agents to understand semantics and
> > subsequently build queries to retrieve data.
> > Most important is probably an execute_query MCP tool.
> >
> >
> >   *
> > Input: data source id, query
> >   *
> > Output
> >      *
> > Data
> >      *
> > Errors
> >      *
> > Suggestions
> >      *
> > Filter value alternatives
> >      *
> > ...
> >
> > This operates under the assumption that the agent is able to discover/see
> > the Ossie semantic model (e.g. using
> http://agenticresourcediscovery.org/).
> > We already have a few variants of execute_query tool across our engines.
> >
> > I'm not sure about query_model (yet), at least for now we see a
> dependency
> > on the agent harness. Some harness can only digest a small view of the
> > semantic model and thus need tools to progressively discover, others
> (like
> > e.g. GitHub copilot cli), don't need any of that, as they'll have full
> > filesystem access, download/query local.
> >
> > ________________________________
> > From: jbonofre (via GitHub) <gi...@apache.org>
> > Sent: Monday, October 5, 2026 8:01 AM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> >
> > GitHub user jbonofre added a comment to the discussion: Proposal: an MCP
> > server for Ossie (and where it should live)
> >
> > I'm not sure we need MCP server at all in Ossie. That could be something
> > catalogs can do.
> >
> > I would rather not consider MCP service directly in Ossie for now (I
> don't
> > see the value).
> >
> > GitHub link:
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> >
> > ----
> > This is an automatically sent email for dev@ossie.apache.org.
> > To unsubscribe, please send an email to:
> dev-unsubscribe@ossie.apache.org
> >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Aleksandr Kosachev <a....@gmail.com>.
Hi Markus, JB,

One gap the thread has not named yet. The REST API spec proposed on 09-23
covers two things: managing models and binding a model to physical
entities. It does not cover running a query and returning data. That is
where execute_query starts. So an MCP spec next to the REST spec would not
duplicate it. It would fill the missing step.

Markus asked how an agent that found a model knows how to run a query
against the source.
In ossie-clickhouse [1], an MCP server over ClickHouse that I maintain, it
does not need to know:
- The model carries no connection.
- The server starts with a model and a database connection.
- The agent sees only the tools.

This reconciles the two positions in the thread. Ossie specifies the tool
contract. The engine or catalog that serves it owns the binding.

One question needs an answer before a spec: what is the input of
execute_query? PR #452 names two query interfaces:
- Layer 2: SQL with MEASURE().
- Layer 3: a declarative query. Metrics, dimensions, filters. No SQL.

execute_query could accept either. ossie-clickhouse takes the Layer 3 form.
When its planner refuses a query, the refusal names the cause.
For example, "unknown metric 'x' (did you mean: ...)", or which metric
would be counted once per joined row.
The agent reads that and repairs the query.
Markus listed four outputs: data, errors, suggestions, filter value
alternatives.
ossie-clickhouse returns data, errors and name suggestions, plus the
generated SQL. It has no filter value alternatives yet.
I would like to hear how those look on your side.

On query_model and the harness dependency, ossie-clickhouse confirms it.
Its overview tool returns datasets and metrics without fields.
Agents then call a search tool and a describe-one-object tool before every
query. A harness with the file on disk, as Markus described, does not need
them. So discovery tools belong in the spec as optional, not dropped.

Warm regards,
Alex Kosachev

[1] https://github.com/Dactopus/ossie-clickhouse
<https://www.google.com/url?q=https://github.com/Dactopus/ossie-clickhouse&source=gmail&ust=1791327988788000&sa=E>

On Tue, Oct 06, 2026 03:08 AM, Markus Cozowicz <m...@cozowicz.at> wrote:

> I'm curious about the expectations here. If e.g. GitHub copilot discovers
> an Ossie semantic model, how would it know how to connect/execute a query
> against the data source (e.g. PBI semantic model, Databricks, Snowflake,
> ...)? Is the expectation that the harness would get additional instructions
> anyway?
>
> Just to re-emphasize, I'm proposing MCP spec along REST API spec, not
> implementation.
> ________________________________
> From: Jean-Baptiste Onofré <jb...@nanthrax.net>
> Sent: Monday, October 5, 2026 8:25 PM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
> Hi
>
> It's not clear to me if we are discussing the specification or the
> implementation. MCP implementation in Ossie: I'm not convinced; it seems
> more like an engine/catalog implementation.
> The spec could make sense.
>
> But do we really need the MCP spec in Ossie for agent model discovery? I'm
> not sure. That's actually one of the reasons for the REST API Spec too
> (model discovery, versioning, etc).
> We could have the MCP spec alongside the REST API Spec.
>
> Also, MCP is good to have but clearly not a requirement for the agent (API
> or CLI are classic data providers for agent, including operations).
>
> Regards
> JB
>
> On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:
>
> > I'd like to provide a slightly different pitch for an MCP server spec
> (not
> > necessarily the implementation). From my perspective, Ossies biggest
> value
> > prop is related to empowering agents to understand semantics and
> > subsequently build queries to retrieve data.
> > Most important is probably an execute_query MCP tool.
> >
> >
> >   *
> > Input: data source id, query
> >   *
> > Output
> >      *
> > Data
> >      *
> > Errors
> >      *
> > Suggestions
> >      *
> > Filter value alternatives
> >      *
> > ...
> >
> > This operates under the assumption that the agent is able to discover/see
> > the Ossie semantic model (e.g. using
> http://agenticresourcediscovery.org/).
> > We already have a few variants of execute_query tool across our engines.
> >
> > I'm not sure about query_model (yet), at least for now we see a
> dependency
> > on the agent harness. Some harness can only digest a small view of the
> > semantic model and thus need tools to progressively discover, others
> (like
> > e.g. GitHub copilot cli), don't need any of that, as they'll have full
> > filesystem access, download/query local.
> >
> > ________________________________
> > From: jbonofre (via GitHub) <gi...@apache.org>
> > Sent: Monday, October 5, 2026 8:01 AM
> > To: dev@ossie.apache.org <de...@ossie.apache.org>
> > Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> > live) [ossie]
> >
> >
> > GitHub user jbonofre added a comment to the discussion: Proposal: an MCP
> > server for Ossie (and where it should live)
> >
> > I'm not sure we need MCP server at all in Ossie. That could be something
> > catalogs can do.
> >
> > I would rather not consider MCP service directly in Ossie for now (I
> don't
> > see the value).
> >
> > GitHub link:
> >
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
> >
> > ----
> > This is an automatically sent email for dev@ossie.apache.org.
> > To unsubscribe, please send an email to:
> dev-unsubscribe@ossie.apache.org
> >
> >
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
I'm curious about the expectations here. If e.g. GitHub copilot discovers an Ossie semantic model, how would it know how to connect/execute a query against the data source (e.g. PBI semantic model, Databricks, Snowflake, ...)? Is the expectation that the harness would get additional instructions anyway?

Just to re-emphasize, I'm proposing MCP spec along REST API spec, not implementation.
________________________________
From: Jean-Baptiste Onofré <jb...@nanthrax.net>
Sent: Monday, October 5, 2026 8:25 PM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Hi

It's not clear to me if we are discussing the specification or the
implementation. MCP implementation in Ossie: I'm not convinced; it seems
more like an engine/catalog implementation.
The spec could make sense.

But do we really need the MCP spec in Ossie for agent model discovery? I'm
not sure. That's actually one of the reasons for the REST API Spec too
(model discovery, versioning, etc).
We could have the MCP spec alongside the REST API Spec.

Also, MCP is good to have but clearly not a requirement for the agent (API
or CLI are classic data providers for agent, including operations).

Regards
JB

On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:

> I'd like to provide a slightly different pitch for an MCP server spec (not
> necessarily the implementation). From my perspective, Ossies biggest value
> prop is related to empowering agents to understand semantics and
> subsequently build queries to retrieve data.
> Most important is probably an execute_query MCP tool.
>
>
>   *
> Input: data source id, query
>   *
> Output
>      *
> Data
>      *
> Errors
>      *
> Suggestions
>      *
> Filter value alternatives
>      *
> ...
>
> This operates under the assumption that the agent is able to discover/see
> the Ossie semantic model (e.g. using http://agenticresourcediscovery.org/).
> We already have a few variants of execute_query tool across our engines.
>
> I'm not sure about query_model (yet), at least for now we see a dependency
> on the agent harness. Some harness can only digest a small view of the
> semantic model and thus need tools to progressively discover, others (like
> e.g. GitHub copilot cli), don't need any of that, as they'll have full
> filesystem access, download/query local.
>
> ________________________________
> From: jbonofre (via GitHub) <gi...@apache.org>
> Sent: Monday, October 5, 2026 8:01 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
>
> GitHub user jbonofre added a comment to the discussion: Proposal: an MCP
> server for Ossie (and where it should live)
>
> I'm not sure we need MCP server at all in Ossie. That could be something
> catalogs can do.
>
> I would rather not consider MCP service directly in Ossie for now (I don't
> see the value).
>
> GitHub link:
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>
> ----
> This is an automatically sent email for dev@ossie.apache.org.
> To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org
>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Jean-Baptiste Onofré <jb...@nanthrax.net>.
Hi

It's not clear to me if we are discussing the specification or the
implementation. MCP implementation in Ossie: I'm not convinced; it seems
more like an engine/catalog implementation.
The spec could make sense.

But do we really need the MCP spec in Ossie for agent model discovery? I'm
not sure. That's actually one of the reasons for the REST API Spec too
(model discovery, versioning, etc).
We could have the MCP spec alongside the REST API Spec.

Also, MCP is good to have but clearly not a requirement for the agent (API
or CLI are classic data providers for agent, including operations).

Regards
JB

On Mon, Oct 5, 2026 at 3:13 PM Markus Cozowicz <m...@cozowicz.at> wrote:

> I'd like to provide a slightly different pitch for an MCP server spec (not
> necessarily the implementation). From my perspective, Ossies biggest value
> prop is related to empowering agents to understand semantics and
> subsequently build queries to retrieve data.
> Most important is probably an execute_query MCP tool.
>
>
>   *
> Input: data source id, query
>   *
> Output
>      *
> Data
>      *
> Errors
>      *
> Suggestions
>      *
> Filter value alternatives
>      *
> ...
>
> This operates under the assumption that the agent is able to discover/see
> the Ossie semantic model (e.g. using http://agenticresourcediscovery.org/).
> We already have a few variants of execute_query tool across our engines.
>
> I'm not sure about query_model (yet), at least for now we see a dependency
> on the agent harness. Some harness can only digest a small view of the
> semantic model and thus need tools to progressively discover, others (like
> e.g. GitHub copilot cli), don't need any of that, as they'll have full
> filesystem access, download/query local.
>
> ________________________________
> From: jbonofre (via GitHub) <gi...@apache.org>
> Sent: Monday, October 5, 2026 8:01 AM
> To: dev@ossie.apache.org <de...@ossie.apache.org>
> Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should
> live) [ossie]
>
>
> GitHub user jbonofre added a comment to the discussion: Proposal: an MCP
> server for Ossie (and where it should live)
>
> I'm not sure we need MCP server at all in Ossie. That could be something
> catalogs can do.
>
> I would rather not consider MCP service directly in Ossie for now (I don't
> see the value).
>
> GitHub link:
> https://github.com/apache/ossie/discussions/510#discussioncomment-18754069
>
> ----
> This is an automatically sent email for dev@ossie.apache.org.
> To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org
>
>

Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by Markus Cozowicz <m...@cozowicz.at>.
I'd like to provide a slightly different pitch for an MCP server spec (not necessarily the implementation). From my perspective, Ossies biggest value prop is related to empowering agents to understand semantics and subsequently build queries to retrieve data.
Most important is probably an execute_query MCP tool.


  *
Input: data source id, query
  *
Output
     *
Data
     *
Errors
     *
Suggestions
     *
Filter value alternatives
     *
...

This operates under the assumption that the agent is able to discover/see the Ossie semantic model (e.g. using http://agenticresourcediscovery.org/). We already have a few variants of execute_query tool across our engines.

I'm not sure about query_model (yet), at least for now we see a dependency on the agent harness. Some harness can only digest a small view of the semantic model and thus need tools to progressively discover, others (like e.g. GitHub copilot cli), don't need any of that, as they'll have full filesystem access, download/query local.

________________________________
From: jbonofre (via GitHub) <gi...@apache.org>
Sent: Monday, October 5, 2026 8:01 AM
To: dev@ossie.apache.org <de...@ossie.apache.org>
Subject: Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]


GitHub user jbonofre added a comment to the discussion: Proposal: an MCP server for Ossie (and where it should live)

I'm not sure we need MCP server at all in Ossie. That could be something catalogs can do.

I would rather not consider MCP service directly in Ossie for now (I don't see the value).

GitHub link: https://github.com/apache/ossie/discussions/510#discussioncomment-18754069

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by "jakub-moravec (via GitHub)" <gi...@apache.org>.
GitHub user jakub-moravec added a comment to the discussion: Proposal: an MCP server for Ossie (and where it should live)

@Ashfaqbs  The distinction of responsibilities for validate (Ossie) and convert/query (vendors) makes a lot of sense to me.

GitHub link: https://github.com/apache/ossie/discussions/510#discussioncomment-18807689

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by "jbonofre (via GitHub)" <gi...@apache.org>.
GitHub user jbonofre added a comment to the discussion: Proposal: an MCP server for Ossie (and where it should live)

I'm not sure we need MCP server at all in Ossie. That could be something catalogs can do. 

I would rather not consider MCP service directly in Ossie for now (I don't see the value).

GitHub link: https://github.com/apache/ossie/discussions/510#discussioncomment-18754069

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by "jakub-moravec (via GitHub)" <gi...@apache.org>.
GitHub user jakub-moravec added a comment to the discussion: Proposal: an MCP server for Ossie (and where it should live)

@Ashfaqbs can you think of any tools that would have to be built out for this? Isn't it sufficient to define skills (potentially as part of Ossie), and then offload the tools/MCP support to vendors?

GitHub link: https://github.com/apache/ossie/discussions/510#discussioncomment-18794586

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org


Re: [D] Proposal: an MCP server for Ossie (and where it should live) [ossie]

Posted by "Ashfaqbs (via GitHub)" <gi...@apache.org>.
GitHub user Ashfaqbs added a comment to the discussion: Proposal: an MCP server for Ossie (and where it should live)

@jakub-moravec good question, and I think it cuts to the actual disagreement with @jbonofre too, so let me try to answer both at once.

A "skill" (in the sense discussion #374 means it, and in the sense most agent frameworks use the word) is advisory: it's text that tells an agent how to think about a problem - "here's the shape of an Ossie model, here's what a metric/relationship looks like, here's how custom_extensions work per vendor." It doesn't give the agent a way to *check* whether what it produced is actually valid. With only a skill, the agent still free-types YAML from memory of the schema, and discussion #374's issue-sprawl problem - agents guessing at the schema and filing issues when they guessed wrong - doesn't go away. A skill changes how good the guess is; it doesn't remove the guessing.

`validate_model` specifically isn't new tooling in the sense of "a thing Ossie has to build and maintain" - it's a thin MCP wrapper around the CLI that already exists in `validation/`. The only thing it adds is making that existing validator callable as a structured round-trip instead of a shell-out-and-parse-stdout step, which is what lets an agent self-correct in a loop rather than producing one guess and stopping. That's the piece I don't think "define skills, offload the rest to vendors" covers: the validator is the project's single source of truth on what's a valid model, and if every vendor wraps it separately, you get N slightly different wrappers around the same schema instead of one.

`list_vendors`/`describe_dialect` are in the same category - they just read `core-spec/spec.yaml`'s existing enum/registry, so there's no new logic to maintain, only a narrower interface to the data that's already there.

Where I think @jbonofre's skepticism is fair: `convert_model` and `query_model` are a different story - those would be real new surface area, and I'm not asking for those to be greenlit here. If there's interest in testing whether this is worth the maintenance cost at all, I'd rather scope a first PR to just `validate_model` (zero new logic, wraps `validation/` as-is) and see if it actually reduces the issue-sprawl discussion #374 describes, before considering the rest.

GitHub link: https://github.com/apache/ossie/discussions/510#discussioncomment-18805036

----
This is an automatically sent email for dev@ossie.apache.org.
To unsubscribe, please send an email to: dev-unsubscribe@ossie.apache.org