feat(mcp): mcp python context propagation #1524

anuraaga · 2025-04-17T05:48:44Z

I am working with @codefromthecrypt on MCP instrumentation for context propagation and he suggested I can send a PR here for it. This adds instrumentation for MCP Python SDK as an implementation of the discussion in modelcontextprotocol/modelcontextprotocol#246.

The key point is picking requst.params._meta as the carrier for context. If a different field was defined in the protocol in the future, it would be relatively simple to migrate to it as the instrumentation otherwise doesn't change at all.

I have not included any examples / quickstart yet since it seemed a bit awkward to me because the instrumentation doesn't generate telemetry itself - it means either having examples that include other instrumentation, or manually creating spans, which both seemed out of scope for a example. In the future if generating MCP telemetry, this problem would be solved. But let me know if you have any advice on this point.

github-actions · 2025-04-17T05:48:58Z

CLA Assistant Lite bot All contributors have signed the CLA ✍️ ✅

anuraaga · 2025-04-17T05:49:48Z

I have read the CLA Document and I hereby sign the CLA

codefromthecrypt · 2025-04-17T05:52:46Z

@anuraaga good point on example.. how about raising a PR on this pointing the requirements.txt to your branch here? It should solve the broken trace. This example already uses openinference for openai-agents and MCP in a way that this should fix, I think.

anuraaga · 2025-04-17T05:52:15Z

python/instrumentation/openinference-instrumentation-mcp/tests/collector.py

+        self.traces.clear()
+
+
+class OTLPServer(HTTPServer):


We need to start up an HTTP collector mostly since the stdio transport starts a subprocess. I considered just returning the parent span context in the tool's response, but figured it's worth setting this up now so it can be extended to actually generating telemetry in the future without new test setup

agree this is the neatest way out I can think of, and also the MCP general discussion has some hints about OTLP proxying on the agent (MCP-client side), so this doubles as a hint of a potential future.

anuraaga · 2025-04-17T05:52:45Z

python/instrumentation/openinference-instrumentation-mcp/tests/test_instrumenter.py

+                await client.initialize()
+                yield client
+        case "sse":
+            proc = await asyncio.create_subprocess_exec(


While the SSE version could start the server in the same process, it seemed simplest to keep it consistent with stdio for these tests

Makes sense

codefromthecrypt · 2025-04-17T05:56:22Z

probably after folks buy-in in general, the next step could be to find where arize examples live and make something like I did, except in typescript for the https://github.com/Arize-ai/phoenix/tree/main/js/packages/phoenix-mcp server. that would be pretty cool

anuraaga · 2025-04-17T07:41:50Z

how about raising a PR on this pointing the requirements.txt to your branch here? It should solve the broken trace.

Good idea. Just one extra requirements dependency fixed the trace

elastic/observability-examples#62

the next step could be to find where arize examples live and make something like I did, except in typescript for the https://github.com/Arize-ai/phoenix/tree/main/js/packages/phoenix-mcp server

Makes sense, will work on Typescript next.

mikeldking

Thanks so much for the PR @anuraaga and the intro @codefromthecrypt ! This seems like a great start to me but will get some folks on my team to take a look just for context building.

cc @axiomofjoy

mikeldking · 2025-04-17T14:48:43Z

...mentation/openinference-instrumentation-mcp/src/openinference/instrumentation/mcp/version.py

@@ -0,0 +1 @@
+__version__ = "1.0.0"


Will also need entries in:

https://github.com/Arize-ai/openinference/blob/main/.release-please-manifest.json
https://github.com/Arize-ai/openinference/blob/main/release-please-config.json

axiomofjoy

This is a solid foundation to build on. Small comments and questions, but it's ready to merge after those are addressed.

High-level question: My understanding is that MCP also allows servers to invoke clients. Will we need to handle context propagation in the opposite direction?

axiomofjoy · 2025-04-17T20:28:37Z

python/tox.ini

  beeai: uv pip install --reinstall {toxinidir}/instrumentation/openinference-instrumentation-beeai[test]
+  mcp: uv pip uninstall -r test-requirements.txt
+  mcp: uv pip install --reinstall-package openinference-instrumentation-mcp .
+  mcp: python -c 'import openinference.instrumentation.mcp'


What is this import needed for?

python/instrumentation/openinference-instrumentation-mcp/pyproject.toml

python/instrumentation/openinference-instrumentation-mcp/test-requirements.txt

mikeldking

excited to try this out! Will merge and publish - just need to get the release-please in place.

This reverts commit 0e78a0b.

axiomofjoy · 2025-04-17T23:37:55Z

It's out: https://pypi.org/project/openinference-instrumentation-mcp/

Keep cooking

anuraaga · 2025-04-18T00:42:52Z

Thanks for the help! Especially 82b3972 sorry I had forgot to fix that minimum version, but you found it.

anuraaga · 2025-04-18T06:22:16Z

High-level question: My understanding is that MCP also allows servers to invoke clients. Will we need to handle context propagation in the opposite direction?

This is a good point - I think the standard tools / resources wrappers on top of the protocol (what MCP SDK calls FastMCP some times) does not expose sending requests from server to client, while the underlying protocol layer indeed supports it. I'll look at getting demo code of that case set up to explore it - I'm a bit scared of how the trace may render if the parent/child get reversed in the middle of a request, I don't think I've ever tried something like that before.

axiomofjoy · 2025-04-18T18:27:43Z

his is a good point - I think the standard tools / resources wrappers on top of the protocol (what MCP SDK calls FastMCP some times) does not expose sending requests from server to client, while the underlying protocol layer indeed supports it. I'll look at getting demo code of that case set up to explore it - I'm a bit scared of how the trace may render if the parent/child get reversed in the middle of a request, I don't think I've ever tried something like that before.

Awesome, excited to see what you find

codefromthecrypt · 2025-04-21T03:54:39Z

For the follow-up,I think the best use case in the MCP protocol to showcase reverse propagation is sampling (server asking a client to do an LLM completion). I elaborated that here and hope it helps! #1543

Co-authored-by: Alexander Song <[email protected]>

feat: mcp python context propagation

6e179ce

anuraaga requested a review from a team as a code owner April 17, 2025 05:48

github-project-automation bot added this to Instrumentation Apr 17, 2025

dosubot bot added the size:L This PR changes 100-499 lines, ignoring generated files. label Apr 17, 2025

Remove accidental lock

17cdbbe

github-actions bot added a commit that referenced this pull request Apr 17, 2025

@anuraaga has signed the CLA in #1524

f6006bf

anuraaga commented Apr 17, 2025

View reviewed changes

mikeldking reviewed Apr 17, 2025

View reviewed changes

axiomofjoy self-requested a review April 17, 2025 20:23

axiomofjoy reviewed Apr 17, 2025

View reviewed changes

mikeldking approved these changes Apr 17, 2025

View reviewed changes

axiomofjoy added 5 commits April 17, 2025 16:15

release please

b0517bb

test version

0e78a0b

add space to pyproject.toml

382abbe

Revert "test version"

cede17f

This reverts commit 0e78a0b.

bump minimum supported version

82b3972

axiomofjoy changed the title ~~feat: mcp python context propagation~~ feat(mcp): mcp python context propagation Apr 17, 2025

axiomofjoy approved these changes Apr 17, 2025

View reviewed changes

axiomofjoy merged commit 1af5f7d into Arize-ai:main Apr 17, 2025
9 checks passed

github-project-automation bot moved this to Done in Instrumentation Apr 17, 2025

github-actions bot mentioned this pull request Apr 17, 2025

chore: release main #1534

Merged

anuraaga mentioned this pull request Apr 18, 2025

feat: mcp js context propagation #1536

Merged

codefromthecrypt mentioned this pull request Apr 21, 2025

[feature request] MCP sampling propagation #1543

Open

anuraaga mentioned this pull request Apr 24, 2025

feat(mcp-js): move instrumentation to transport layer #1555

Merged

github-actions bot mentioned this pull request Apr 24, 2025

chore(js): update versions #1558

Merged

GeLi2001 pushed a commit that referenced this pull request May 28, 2025

feat(mcp): mcp python context propagation (#1524)

ed8053c

Co-authored-by: Alexander Song <[email protected]>

feat(mcp): mcp python context propagation #1524

feat(mcp): mcp python context propagation #1524

Uh oh!

Conversation

anuraaga commented Apr 17, 2025

Uh oh!

github-actions bot commented Apr 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

anuraaga commented Apr 17, 2025

Uh oh!

codefromthecrypt commented Apr 17, 2025

Uh oh!

anuraaga Apr 17, 2025

Choose a reason for hiding this comment

Uh oh!

codefromthecrypt Apr 17, 2025

Choose a reason for hiding this comment

Uh oh!

anuraaga Apr 17, 2025

Choose a reason for hiding this comment

Uh oh!

axiomofjoy Apr 17, 2025

Choose a reason for hiding this comment

Uh oh!

codefromthecrypt commented Apr 17, 2025

Uh oh!

anuraaga commented Apr 17, 2025

Uh oh!

mikeldking left a comment

Choose a reason for hiding this comment

Uh oh!

mikeldking Apr 17, 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!

axiomofjoy 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!

axiomofjoy Apr 17, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

mikeldking left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

axiomofjoy commented Apr 17, 2025

Uh oh!

anuraaga commented Apr 18, 2025

Uh oh!

anuraaga commented Apr 18, 2025

Uh oh!

axiomofjoy commented Apr 18, 2025

Uh oh!

codefromthecrypt commented Apr 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions bot commented Apr 17, 2025 •

edited

Loading

mikeldking Apr 17, 2025 •

edited

Loading

axiomofjoy left a comment •

edited

Loading