madeye/mcp-distill

mcp-distill is a recording server for AI coding tools that captures every prompt, context snippet, and response from Claude and Codex so you can use those interactions as training data to fine-tune a smaller model. The idea behind small-model distillation is that the responses from a large, expensive model can teach a smaller, cheaper model to behave similarly, but first you need the raw conversation data. The server speaks the Model Context Protocol (MCP), a standard way for AI coding assistants to communicate with external tools. Once installed alongside Codex or Claude Code, it automatically intercepts each conversation turn and saves it losslessly as a JSONL file, a format where each line is a self-contained JSON record. The data is organized by date under a configurable storage directory. When you want to train a model, you export the collected sessions in one of three formats: openai_chat, sharegpt, or anthropic, depending on which training framework you are using. Installation is handled by the tool itself. Running the install command for either Codex or Claude Code modifies the relevant configuration file to add mcp-distill as a registered server. The recording starts automatically when the server launches, so no manual setup is needed for routine use. You can pause and resume recording, start named sessions, append turns manually, or export a dataset on demand through exposed MCP tools. The server is built in Rust and compiled to a single binary. It supports optional zstd compression for stored sessions. The export output can be fed directly into training frameworks like Unsloth with minimal preprocessing.