wtf is tracekit?

tracekit is a command line tool that takes the logs your coding agent already creates, such as Claude Code's session history, and turns them into a small language model trained on your own habits. Every time you use a coding agent it writes detailed records of your edits and commit messages, and these records normally just sit unused on your computer. tracekit reads that history, keeps only the parts that were actually kept and committed, and fine tunes a small model on them. The resulting model is meant to run entirely on your own machine, for routine tasks like writing commit messages or making small edits, in a style that matches how you normally work. It is not meant to replace a large model like Claude for complex reasoning. Instead it aims to take over a portion of your simpler, repetitive requests so that fewer of them need to go through a paid API. The tool works as a pipeline of steps you can run individually or all together with one command. First it scans your machine for existing coding agent logs, then it parses them into a local database, filters out edits that were reverted or unused, builds a training dataset from what is left, fine tunes a small model using a technique called QLoRA, evaluates the result, and finally exports it to run with Ollama, a tool for running language models locally. Everything happens on your own computer. The README states there are no network calls in the core product and no telemetry. Secrets found in your logs are automatically filtered out before anything is stored, and file contents are never kept, only a short fingerprint of them is saved instead. Right now it supports reading logs from Claude Code, with support for other coding agents like Codex, Aider, and Cursor listed as planned but not yet available. Training requires Python 3.11 or later, and a GPU is recommended for the training step, though a slower CPU option exists for testing purposes.