traceroot-ai/traceroot

TraceRoot – Self‑Improving Layer for AI Coding Agents

What it is – TraceRoot is an open‑source platform that records ("traces") every call an AI coding agent makes to a language model or tool, lets you define detectors that flag undesirable behaviour, clusters those flags into signals, and then runs automated evals to verify fixes. The idea is to close the loop: production failures become concrete improvement suggestions that a coding‑agent can act on, test, and redeploy.

Key components

Feature	What you get
Trace	Full‑stack observability of model calls, tool calls, latency, cost, and inputs/outputs.
Detect	Customisable detectors that encode “good behaviour” and automatically flag runs that violate it.
Signals	Clustering of detector outputs; a background coding‑agent inspects the related code (via GitHub sandbox) and proposes fixes.
Verify	Offline evaluation against your own datasets to measure improvements and catch regressions.
CLI / SDKs	Agent‑native CLI plus Python (traceroot-py) and TypeScript (traceroot-ts) SDKs for easy instrumentation.
Integrations	Ready‑made hooks for major agent frameworks (LangChain, CrewAI, AutoGen, etc.) and model providers (OpenAI, Anthropic, Gemini, Mistral, OpenRouter).

How it works

1. Instrument your application with the TraceRoot SDK or let a coding‑agent set it up via the CLI.
2. Run your AI‑powered code in production; each LLM/tool call is sent to the TraceRoot backend.
3. Detect – configured detectors automatically label runs that break your expectations.
4. Signal – flagged runs are grouped; a background agent examines the trace and the associated repository code, then suggests a code change.
5. Verify – the suggested change is tested with offline evals before being merged, completing the self‑improvement cycle.

Getting started

• Cloud – Sign up at app.traceroot.ai for a hosted instance.
• Self‑host – Clone the repo, copy .env.example → .env, and run make prod-lite (Docker‑Compose). Access the UI at localhost:3000.
• Instrumentation – Install the CLI (npm i -g traceroot-cli) and run traceroot login. Then ask your coding agent to set up tracing, or manually add the Python/TS SDK and call TraceRoot.initialize(...).

Who it’s for

• Teams building AI‑assisted coding assistants or other LLM‑driven agents that need production‑grade observability and a systematic way to turn bugs into verified fixes.
• Developers who already use frameworks like LangChain, CrewAI, or LlamaIndex and want a plug‑and‑play monitoring layer.

License – Apache 2.0 for the core code; some ee/ directories are under a commercial Enterprise License.

Community – Active Discord, Y Combinator backing, and contribution guides for newcomers.

---

All details are taken directly from the repository’s README.