jupyter-mcp-spaces

<!-- ~ Copyright (c) 2024- Datalayer, Inc. ~ ~ BSD 3-Clause License --> # Jupyter MCP Spaces Lists and opens the notebooks of a [Datalayer](https://datalayer.ai) space. Datalayer keeps notebooks in **spaces**, addressed by uid. There is no filesystem and no kernels API, so the tools that assume a Jupyter server have nothing to answer: | Tool | On a Jupyter server | On Datalayer | |---|---|---| | `list_files` | lists the directory | nothing to list | | `list_kernels` | lists kernels | runtimes, provisioned on demand | | `list_notebooks` | notebooks bound this session | — | This extension replaces `list_notebooks` with one that answers the question people actually ask, adds `list_spaces` and `find_notebook`, and hides the tools that cannot work — so an agent is never offered a tool that always fails. ## Install ```bash pip install jupyter-mcp-spaces ``` It registers itself on the `jupyter_mcp_server.extensions` entry point and activates only when the server runs with `--document-provider datalayer`. Pointed at a Jupyter server it does nothing, and the ordinary tools stay. ```bash jupyter-mcp-server start --transport streamable-http \ --document-provider datalayer \ --document-url https://prod1.datalayer.run ``` ## Tools | Tool | What it answers | |---|---| | `list_spaces` | The spaces you can reach | | `list_notebooks` | The notebooks in them, with uid and space | | `find_notebook` | Which notebook a name means — or the candidates, when it is ambiguous | `find_notebook` never picks between candidates. Guessing which notebook somebody meant is how an agent edits the wrong one. ## Credentials Every call carries the token of the request being served, from `jupyter_mcp_server.identity`. A server acting for several people over its life must use the credential each request arrived with, not one configured at startup — see the [identity documentation](https://jupyter-mcp-server.datalayer.tech/features/identity).