ModelContextProtocol MCP server for AI agentic collaboration

[complyue]

Problem

Besides human collaborations with the lab, it's very desirable for AI agents easily get in the loop. ModelContextProtocol (https://modelcontextprotocol.io) makes this idea very practical.

Proposed Solution

Ideally an sse endpoint (https://modelcontextprotocol.io/specification/2025-06-18/basic/transports#streamable-http) of MCP can be started with jupyterlab server, embedded as a plugin, for AI tools to easily config their MCP access to a jupyterlab instance with mere URL.

Or if notebook editing (and possibly other) apis can be exposed via typescript sdk of juypterlab, AI tools can easily config the MCP access with simple npx cmdls.

Additional context

@jupyterlab/services works now in accessing jupyterlab notebooks and other stuff, but lacks RTC ability for user to monitor and see AI mods in realtime.

[complyue]

https://github.com/datalayer/jupyter-mcp-server claims to have a working implementation, but I'm quite unhappy with their setup approach:

pip uninstall -y pycrdt datalayer_pycrdt
pip install datalayer_pycrdt==0.12.17

[krassowski]

Also related:

• [v3-beta] Integrate with MCP jupyter-ai#1352

[krassowski]

And:

• https://github.com/jupyter-ai-contrib/jupyter-server-ai-tools

[davidbrochart]

https://github.com/datalayer/jupyter-mcp-server claims to have a working implementation, but I'm quite unhappy with their setup approach:

pip uninstall -y pycrdt datalayer_pycrdt
pip install datalayer_pycrdt==0.12.17

Indeed, I wouldn't trust a closed-source fork of pycrdt.

[complyue]

drafting an implementation at https://github.com/complyue/jupyter-collaboration-mcp , at least list_notebooks is working now.

---

update: unable to gather enough backend api to support all functionalities there, drafting a typescript based frontend implementation at https://github.com/complyue/jupyterlab-rtc-mcp

[complyue]

Now the ts based MCP server impl -- https://github.com/complyue/jupyterlab-rtc-mcp -- can sorta be considered alpha quality, would anyone like to try it and see if we can make its way to be accepted somewhere for collaborative maintenance and published to npm registry?

I haven't get programatic tests in the repo so far, instead I have AI agent made the plan and performed validation tests with configured MCP server:

JupyterLabRTC-MCP-validations.md