FAQ

Which browsers and agents does this work with?

Any browser that exposes document.modelContext and any agent that can call WebMCP tools. It has been verified on the live demo in Chrome with WebMCP enabled, and the project was built for ChatGPT's in-app browser. As of writing no browser ships WebMCP by default, so the extension detects its absence and stays out of the way. The notebook works either way.

Do I need a server, API keys or a chat UI?

No. It's a frontend extension with no server component, no keys and no LLM of its own. Bring your own agent.

Does it work in JupyterLab and Notebook 7, or only JupyterLite?

All three. It's one prebuilt extension, verified against a JupyterLite site (Pyodide kernel), a JupyterLab 4.6 server and a Notebook 7 server, with no platform-specific code. See Install.

Can the agent run arbitrary code?

Only by inserting a visible cell and running it, which you can watch. There is no hidden execution path. See the security model for what that does and doesn't protect against.

What if I edit a cell while the agent is working?

The agent's write is refused with STALE_CELL and your text is untouched. The human always wins.

Can I require approval for each edit?

Yes: switch the Agent panel to Propose mode. It currently covers jupyter_update_cell; insert, delete and run still apply directly (see the roadmap).

Does it work with real-time collaboration?

The agent writes to the same Yjs model that jupyter-collaboration syncs, so remote humans see its edits arrive. The agent doesn't yet appear in the awareness layer as a labelled cursor. Details in Multiplayer.

Is it stable?

It's 0.x. The tool surface may still change between minor versions; the changelog lists changes per release.

How do I report a bug or ask for something?

Open an issue for bugs, or start a thread in Discussions for ideas and questions.