Tool to install dependencies

[Omoeba]

In addition to the run_python_code tool, an additional tool, such as install_python_dependencies would be extremely helpful. Currently, when used as an MCP server, dependencies can only be specified in the client's mcp.json file using --deps. The user has to either choose an all-encompassing set of dependencies that leads to unnecessary bloat for most chats, or have to edit mcp.json and restart the mcp-run-python server when a particular chat needs additional dependencies. With the new tool, a minimal set of dependencies can be specified in mcp.json and additional requirements can be dynamically installed at runtime.

Workflow preserving read-only node_modules

1. Model calls run_python_code and fails due to the lack of a dependency
2. Model calls install_python_dependencies with a list of dependencies to install
3. The existing read-only deno runtime is stopped
4. The specified dependencies are installed using deno with write permissions
5. A new deno runtime is started in read-only mode with the new dependencies
6. The model retries run_python_code, which succeeds with the new dependencies

[DouweM]

@Omoeba I believe installing arbitrary PyPI packages would be unsafe and circumvent the security model described in https://github.com/pydantic/mcp-run-python#dependencies, as installation requires filesystem access, and an arbitrary package could have been uploaded to PyPI by an attacker. So you'd need a package whitelist, but then it's unclear what the advantage would be of that over just including them all by default.

[Omoeba]

@DouweM One advantage I can think of is startup time. Using a package list of numpy, pandas, scipy, matplotlib, scikit-learn, scikit-image, sympy, requests, beautifulsoup4, pillow, opencv-python, lxml, tqdm, pyyaml, pytest results in a startup time of more than 10 seconds. Having a whitelist and dynamically loading will reduce that startup time by a lot.