openinterpretability/openinterp-mcp

openinterp-mcp is a Python toolkit for researchers studying mechanistic interpretability, the field that tries to understand what is actually happening inside AI language models when they process information, rather than treating the model as a black box. The toolkit exposes eight tool primitives that an AI coding assistant, such as Claude Code, Cursor, Cline, or any tool that speaks the MCP protocol, can call to run experiments. The primitives cover: attaching to a running compute session, checking its health, listing loaded probes, extracting activations (the internal numerical signals a model produces at specific layers), evaluating a probe against a stored capture, steering the model's behavior by injecting a direction into its internals, looking up SAE features from a stored activation, and running a standardized causality protocol. That protocol checks whether an observed signal is genuinely causing a behavior or only correlated with it, and returns one of five verdicts, including causal, weak causal, or an epiphenomenal category. Two additional Python modules, not exposed as MCP tools, handle publishing results to a shared registry and running replication checks. The architecture is built to be privacy first. The MCP server runs as a stateless process on the researcher's own laptop, while the actual model runs on the researcher's own compute, for example Google Colab, vast.ai, or runpod, and is reached through an ngrok or cloudflared tunnel. No inference happens on the project's own servers, no API keys are collected, and no queries pass through the project's infrastructure. Setup involves running one cell in a Colab notebook to install the package and launch a session, then connecting an MCP-compatible coding assistant to that session's URL. The toolkit is still early, described as v0.1.0 beta, with the project noting the API may change before a 1.0 release. It is released under the Apache-2.0 license.