skandesh/tracebench

Tracebench is a local tool that lets you browse and review the session logs produced by AI coding assistants. If you have used Claude Code, Codex, or Cursor on your computer, each of those tools quietly writes a log of everything it did during a session: which files it read, what commands it ran, how many tokens it consumed. Tracebench picks up those logs and shows them in a single web-based viewer that runs entirely on your own machine, with no data sent anywhere. You start it with one command, npx tracebench, and it opens a browser tab at a local address. On first launch it scans the standard locations where Claude Code, Codex, and Cursor store their logs, loads everything into a small local database, and presents a three-pane interface. You can see a timeline of every action the agent took, browse sessions across different tools side by side, and review cost estimates based on token counts. The project is built around the idea that most people use more than one AI coding tool, yet existing log viewers are each tied to a single tool. Tracebench introduces a common format that all supported agents convert into, so the same timeline view works regardless of which assistant produced the session. Adding support for a new tool means writing one adapter module that converts its log format into that shared format. Under the hood the server is a lightweight Node.js process using Fastify, storing indexed sessions in SQLite. The front end is built with Vite and React. The codebase is split into six packages: a core library with the shared schema and pricing calculations, one adapter per supported agent, the server, and the UI. The project is at an early version and marks certain adapters as planned rather than available. The roadmap includes context window analysis, cost dashboards, and a plugin system for community-built adapters. The source is released under the Apache 2.0 license.