ssreeni1/tracebase

Tracebase is a tool that captures and inspects logs from AI coding agent sessions, specifically Codex and Claude Code, so a developer can see exactly what an agent run did after the fact. It runs entirely on your own computer: there is no network interception, no hidden capture of the model's private reasoning, and no remote service involved. The tool imports the JSONL transcript files that Codex and Claude Code already write locally, encrypts the raw event data at rest using AES 256 encryption, builds a searchable local index, and serves a dashboard on your machine so you can browse sessions, individual events, and summaries. It can also accept live events sent to it while an agent is running, if you opt into that. Tracebase is built to answer practical questions about an agent run: whether tests actually executed, where the agent got stuck or looped, how many tokens and how much cost were used, which files and tools were touched, whether an exported bundle is safe to share with someone else, and whether one run improved on a previous one. It flags things like repeated searches, repeated commands, and unusually large outputs as signs that the agent may have wasted context without making progress, while acknowledging that some expensive steps, like full UI testing, are sometimes genuinely necessary. By default, everything is stored under a local folder, bound to localhost only, with raw data encrypted and export bundles redacted unless you explicitly ask for a raw export. Hidden model reasoning is never captured unless the underlying provider exposes it itself. Installation is done through npm: installing the tracebase-local package globally gives you commands to initialize, import existing transcripts, and start the local dashboard, which then opens in a browser at a local address. Additional commands support comparing runs, generating summaries, checking capture health, and running an MCP server for integration with other tools. The project is released under the MIT license.