[BugFix] MCP - expose correct header for session id #7160

MagnusS0 · 2025-07-03T22:17:58Z

Fix MCP Session ID

Description

In the MCP PR #7094 I set the stateless_http=True, after some testing I discovered that this caused context length issues with Cursor and the Claud Desktop Client.
This fix reverts to stateless_http=False and fix the previous Session ID bug by exposing the Mcp-Session-Id header.

@piiq @deeleeramone

deeleeramone · 2025-07-04T03:46:18Z

@MagnusS0, can you please add some instructions-for-dummies for connecting and configuring for immediate use in both Claude Desktop and Cursor? We need explicit steps that anyone can copy and paste, thanks.

The update casued problems with converting tools from OpenAPI specs

MagnusS0 · 2025-07-06T12:48:20Z

In terms of how to set it up I added some instructions.

But also notice that if you run it with, openbb[all] all the technical and quantitative endpoints also become tools. I don't think this is something we want, since moste of those endpoints expect inputs as the output of equity_price_historical. This should probably be programmatically passed instead of having the LLM as a layer between.

Potentially converting these tools to automatically fetch the right data would make more sense from an agent perspective.

But maybe just ignore these as standard setting for now is the best approach?

deeleeramone · 2025-07-06T18:33:15Z

Potentially converting these tools to automatically fetch the right data would make more sense from an agent perspective.

I don't think this will work out reliably, there are too many factors to assume the "right" data is being collected and processed. The functions are intended to be used with any time series.

But maybe just ignore these as standard setting for now is the best approach?

This would be fine, IMO, to ignore all POST requests. If there is a clear path towards enabling these via a code interpreter agent, or similar, then we can look at incorporating some. Alternatively, one could manually configure tools that utilize these endpoints via specific instructions, but that's a rabbit hole and out-of-scope for now.

MagnusS0 · 2025-07-06T18:51:29Z

This would be fine, IMO, to ignore all POST requests.

On board with that! Let me add POST to be excluded by default.

Alternatively, one could manually configure tools that utilize these endpoints via specific instructions, but that's a rabbit hole and out-of-scope for now.

I love a good rabbit hole, but maybe someone comes up with a clever way to use them later on 😄

DidierRLopes · 2025-07-08T02:13:14Z

Having issues when running this one :/

deeleeramone · 2025-07-08T03:29:24Z

openbb_platform/extensions/mcp_server/pyproject.toml

@@ -16,7 +16,7 @@ openbb-mcp = "openbb_mcp_server.main:main"
 [tool.poetry.dependencies]
 python = ">=3.10,<3.13"
 openbb-core = "^1.4.3"
-fastmcp = "^2.8.0"
+fastmcp = ">=2.8.0,<2.9.0"


We should avoid pinning the dependency and always favor the latest version. The current version is 2.10.2, so this is already obsolete.

Also, no need to update the version number of the package, that will be handled when the extension is published.

This is due to tool descriptions/responses bloated after 2.9.

I have traced it back to this PR
jlowin/fastmcp#897
We end up with out of context length just reading in a single tool.
Others reported similar issues.
It looks like a fix might have been proposed here jlowin/fastmcp#1092

So temporary fix was to fallback to 2.8.1

This is due to tool descriptions/responses bloated after 2.9.

The PR you are referencing mentions, and is before 2.9:

...this resulted in a test failure and also bloated tool definitions, so I added a fix for that.

This PR has been merged, and was included in 2.9, which suggests the bloated descriptions issue you are referencing is an issue pre-2.9, along with $ref handling, which our openapi.json has in every endpoint.

We need to be working with the latest versions for a number of reasons, but by not using the latest version, we create problems for the immediate future.

jlowin/fastmcp#887

Totally agree on using the newest version, but currently this will make the MCP server unusable.
I think this needs to be fixed in the FastMCP library, if not, we would have to look into how the openapi.json is structured and if we can adjust the customize components function in some way.

you are referencing is an issue pre-2.9

What I meant was that that PR was supposed to fix bloating but for conversion of the OpenBB API specifically it caused bloating, and broke the MCP server.

Eg. here is the news_world tool with v2.8.1

[debug] [Transport] Received: {"jsonrpc":"2.0","id":1,"result":{"tools":[{"name":"news_world","description":"World News. Global news data.","inputSchema":{"type":"object","properties":{"provider":{"type":"string","enum":["benzinga","biztoc","fmp","intrinio","tiingo"],"title":"Provider"},"limit":{"type":"integer","title":"Limit","description":"The number of data entries to return. The number of articles to return.","default":2500},"start_date":{"anyOf":[{"type":"string","format":"date"},{"type":"null"}],"title":"Start Date","description":"Start date of the data, in YYYY-MM-DD format."},"end_date":{"anyOf":[{"type":"string","format":"date"},{"type":"null"}],"title":"End Date","description":"End date of the data, in YYYY-MM-DD format."},"date":{"anyOf":[{"type":"string","format":"date"},{"type":"null"}],"title":"benzinga","description":"A specific date to get data for. (provider: benzinga)"},"display":{"type":"string","enum":["headline","abstract","full"],"title":"benzinga","description":"Specify headline only (headline), headline + teaser (abstract), or headline + full body (full). (provider: benzinga)","default":"full"},"updated_since":{"anyOf":[{"type":"integer"},{"type":"null"}],"title":"benzinga","description":"Number of seconds since the news was updated. (provider: benzinga)"},"published_since":{"anyOf":[{"type":"integer"},{"type":"null"}],"title":"benzinga","description":"Number of seconds since the news was published. (provider: benzinga)"},"sort":{"type":"string","enum":["id","created","updated"],"title":"benzinga","description":"Key to sort the news by. (provider: benzinga)","default":"created"},"order":{"type":"string","enum":["asc","desc"],"title":"benzinga","description":"Order to sort the news by. (provider: benzinga)","default":"desc"},"isin":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"The ISIN of the news to retrieve. (provider: benzinga)"},"cusip":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"The CUSIP of the news to retrieve. (provider: benzinga)"},"channels":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"Channels of the news to retrieve. (provider: benzinga)"},"topics":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"Topics of the news to retrieve. (provider: benzinga)"},"authors":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"Authors of the news to retrieve. (provider: benzinga)"},"content_types":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"benzinga","description":"Content types of the news to retrieve. (provider: benzinga)"},"term":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"biztoc","description":"Search term to filter articles by. This overrides all other filters. (provider: biztoc)"},"source":{"anyOf":[{"type":"string"},{"type":"string","enum":["yahoo","moody","moody_us_news","moody_us_press_releases"]},{"type":"null"}],"title":"biztoc,intrinio,tiingo","description":"Filter by a specific publisher. Only valid when filter is set to source. (provider: biztoc);\n The source of the news article. (provider: intrinio);\n A comma-separated list of the domains requested. Multiple comma separated items allowed. (provider: tiingo)","tiingo":{"multiple_items_allowed":true}},"sentiment":{"anyOf":[{"type":"string","enum":["positive","neutral","negative"]},{"type":"null"}],"title":"intrinio","description":"Return news only from this source. (provider: intrinio)"},"language":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"intrinio","description":"Filter by language. Unsupported for yahoo source. (provider: intrinio)"},"topic":{"anyOf":[{"type":"string"},{"type":"null"}],"title":"intrinio","description":"Filter by topic. Unsupported for yahoo source. (provider: intrinio)"},"word_count_greater_than":{"anyOf":[{"type":"integer"},{"type":"null"}],"title":"intrinio","description":"News stories will have a word count greater than this value. Unsupported for yahoo source. (provider: intrinio)"},"word_count_less_than":{"anyOf":[{"type":"integer"},{"type":"null"}],"title":"intrinio","description":"News stories will have a word count less than this value. Unsupported for yahoo source. (provider: intrinio)"},"is_spam":{"anyOf":[{"type":"boolean"},{"type":"null"}],"title":"intrinio","description":"Filter whether it is marked as spam or not. Unsupported for yahoo source. (provider: intrinio)"},"business_relevance_greater_than":{"anyOf":[{"type":"number"},{"type":"null"}],"title":"intrinio","description":"News stories will have a business relevance score more than this value. Unsupported for yahoo source. Value is a decimal between 0 and 1. (provider: intrinio)"},"business_relevance_less_than":{"anyOf":[{"type":"number"},{"type":"null"}],"title":"intrinio","description":"News stories will have a business relevance score less than this value. Unsupported for yahoo source. Value is a decimal between 0 and 1. (provider: intrinio)"},"offset":{"anyOf":[{"type":"integer"},{"type":"null"}],"title":"tiingo","description":"Page offset, used in conjunction with limit. (provider: tiingo)","default":0}},"required":["provider"]}}

With v2.9 and 2.10 this becomes a close to 1 MB txt file news_tool.txt. The pruning seems to fail so where the example above stops it continues:

"$defs":{"ADFTestModel":{

I'll look into the openapi.json to see if I can get a better understanding.

What @MagnusS0 is saying makes sense. We should pin dependency to 2.8 until it gets fixed upstream, and then bump it to the first version where this works. Particularly if this is going to be addressed, no reason to work on a workaround if it gets fixed at the root.

Also, I was able to run this with uvx well :)

Cloned fastmcp to figure out where it goes wrong so I might be able to create an Issue.
But the main issue is that all schema's are added to every tool.

https://github.com/jlowin/fastmcp/blob/main/src/fastmcp/utilities/json_schema.py

@deeleeramone you probably know the API better than me, are you able to see where the pruning might go wrong?

This has been resolved with the latest commits and we can use the newest version of fastmcp.
Currently there is a output validation error, but since my fix was merged jlowin/fastmcp#1119 from the next release this should be good.

We can wait to merge this until then or set outputs toNone temporarily.

Hi all -- just released 2.10.5 including @MagnusS0 fix in jlowin/fastmcp#1119. Sorry for the trouble! Curious how we can do more to avoid the docstrings getting too long -- the line between "automatically regurgitating all the OpenAPI data" and "not overwhelming the LLM" is a very fine one!

Hey @jlowin 👋 thanks for the quick release and for jumping into the discussion!

One suggestion: the OpenBB Platform MCP that @MagnusS0 is working on could serve as a good test case for experimenting with ways to trim down the OpenAPI docstrings. It’s a bit of an edge case: the API exposes 300+ endpoints by default, and each one is very thoroughly documented. Every parameter and return type is typed and described in detail using Pydantic models. So yeah, lots of context.

A couple of things I noticed in our openapi.json that might be relevant here:

Some define the same info in both the parameter specs and the schema.

2. Many return schemas link to large Pydantic models, and we use anyOf to describe unions of similar data models (we standardize across data providers). This can add a lot of weight to the spec, especially when models are overlapping a lot.
I don’t have a full picture of how MCP uses these specs under the hood (whether the LLM relies on output schemas or not). But if not, maybe skipping the return models could reduce some of the overload? Not proposing that outright, just hoping to spark some discussion. Curious what you and @MagnusS0 think.

Hey @jlowin, thanks for all your awesome work with FastMCP!

Totally agree, finding the sweet spot for docstrings is tricky. Really like the mcp_component_fn adds a lot of flexibility to customize things and not just copy-paste from OpenAPI.

And to @piiq's point, this is spot on, the return model is not information the LLM needs. I actually already handle this with a simple regex that removes them from tool descriptions.

Btw part of the reason this conversion straight from OpenAPI works is because of how great things are documented, much better than missing context!

deeleeramone

Are we good to update the definition to the latest fastmcp version? The pyproject.toml is keeping us at 2.8

MagnusS0 · 2025-07-12T21:36:38Z

Should be ready to merge now 😃

deeleeramone

Thank you very much for your contribution, as well as the cross-library collaboration! Open-source at its finest!

[BugFix] MCP - expose correct header for session id

7dbfd54

MagnusS0 and others added 3 commits July 4, 2025 19:13

Pin FastMCP to <2.9

1fa079e

The update casued problems with converting tools from OpenAPI specs

Update default settings

7fadcb2

Merge branch 'develop' into hotfix/mcp-server-session-fix

70dadd2

MagnusS0 and others added 2 commits July 7, 2025 01:18

Add exclusion for all POST requests in route filtering

3f6f075

Merge branch 'develop' into hotfix/mcp-server-session-fix

7f45fb3

deeleeramone requested changes Jul 8, 2025

View reviewed changes

DidierRLopes and others added 6 commits July 9, 2025 00:32

Merge branch 'develop' into hotfix/mcp-server-session-fix

79b41c6

fix compress input and output schemas with newer fastmcp versions

1118a52

Add unit tests for MCP server, utilities and settings

7bd887a

add missing init file

3c5c354

skip tests for python 3.9

97d2a2c

Add setup instructions for VS Code

6ae6212

deeleeramone reviewed Jul 12, 2025

View reviewed changes

update fastmcp version and lock file

8263590

deeleeramone approved these changes Jul 12, 2025

View reviewed changes

deeleeramone added this pull request to the merge queue Jul 12, 2025

Merged via the queue into OpenBB-finance:develop with commit 7b520e5 Jul 12, 2025
9 checks passed

MagnusS0 deleted the hotfix/mcp-server-session-fix branch July 12, 2025 23:35

[BugFix] MCP - expose correct header for session id #7160

[BugFix] MCP - expose correct header for session id #7160

Uh oh!

Conversation

MagnusS0 commented Jul 3, 2025

Fix MCP Session ID

Description

Uh oh!

deeleeramone commented Jul 4, 2025

Uh oh!

MagnusS0 commented Jul 6, 2025

Uh oh!

deeleeramone commented Jul 6, 2025

Uh oh!

MagnusS0 commented Jul 6, 2025

Uh oh!

DidierRLopes commented Jul 8, 2025

Uh oh!

deeleeramone Jul 8, 2025

Choose a reason for hiding this comment

Uh oh!

MagnusS0 Jul 8, 2025

Choose a reason for hiding this comment

Uh oh!

deeleeramone Jul 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MagnusS0 Jul 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DidierRLopes Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

MagnusS0 Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

MagnusS0 Jul 11, 2025

Choose a reason for hiding this comment

Uh oh!

jlowin Jul 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

piiq Jul 12, 2025

Choose a reason for hiding this comment

Uh oh!

MagnusS0 Jul 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

deeleeramone left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MagnusS0 commented Jul 12, 2025

Uh oh!

deeleeramone left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

deeleeramone Jul 8, 2025 •

edited

Loading

MagnusS0 Jul 8, 2025 •

edited

Loading

jlowin Jul 11, 2025 •

edited

Loading

MagnusS0 Jul 12, 2025 •

edited

Loading

deeleeramone left a comment •

edited

Loading