mcp_run_isolated_python

# MCP Server for Sandboxed Python Code Execution This MCP server allows your LLM to execute python code securely and returns the results - including files. The code is executed in a configurable sandboxed environment, with strong defaults - like no network access and heavily limited file write permissions. It's great for use cases complicated use cases where LLMs run into hallucination. For example, stuff that requires a lot of math - where LLMs are notoriously finicky - or if you want to generate cool graphs. If you this, your LLM will always be able to count the number of "r" in strawberry! ## Quick Start ### MCP - via Docker (recommended) For this to run we have to set some security options on the container. This is due to bubblewrap needing to create user, mount and pid namespaces, which the container has to be allowed to do: - seccomp=unconfined - apparmor=unconfined - systempaths=unconfined This is mandatory - to my knowledge these are the miniumum needed permissions, but feel free to experiment :) ``` docker run -p 6400:6400 \ --pids-limit 512 --memory 2g --cpus 2 \ --security-opt seccomp=unconfined \ --security-opt apparmor=unconfined \ --security-opt systempaths=unconfined \ -e PYTHON_DEPENDENCIES="pydantic numpy" \ kigstn/mcp-run-isolated-python ``` You can pass your CLI settings directly after that, the dockerfile uses entrypoint to start the server and listens to all args. Note: - If you are running on Ubuntu 24 you will need to adapt your apparmor profile. We have a script, just run `setup_host.sh`. [More Info](https://github.com/anthropics/sandbox-runtime/blob/main/.github/workflows/integration-tests.yml#L152) - These three options only loosen the *container*. The code your LLM runs is still sandboxed by `srt` inside it - no network, no writes outside its temp dir, own pid namespace. - Docker automatically creates a separate UV python interpreter for the runtime - so you dont have to pass that :) - To control your packages, use…