traceroot

TraceRoot is designed to address common pain points in scaling production AI agent systems, where manual review of every trace is unsustainable, cross-stack debugging of agent failures is fragmented, and agent improvement processes are often ad-hoc. Its core capabilities include OpenTelemetry-compatible tracing via Python and TypeScript SDKs, which captures full LLM calls, tool use, and multi-step agent runs, with visibility into inputs, outputs, latency, token consumption, and associated costs. It offers configurable detectors that automatically screen incoming production traces to flag hallucinations, tool failures, logic errors, and safety issues, with adjustable sampling rates and selectable judge models for review. The platform includes dataset tooling to author, version, and publish test cases that agents are expected to handle reliably, plus offline evaluation workflows that run agent systems against curated datasets, provide per-case scoring, and compare performance across candidate versions, with every test case stored as an inspectable trace. A dedicated CLI tool lets users read, export, and inspect traces and detector findings to integrate agent debugging into existing coding workflows. Built-in dashboards and configurable threshold alerts track agent quality, latency, and cost over time, and an in-app AI assistant can explore trace data with access to connected source code and GitHub context, supporting either hosted models or user-provided API keys. The platform includes pre-built instrumentation for a wide range of popular agent frameworks, including LangChain and LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Claude Agent SDK, LlamaIndex, DSPy, Pydantic AI, Mastra, and Vercel AI SDK, as well as native tracing support for major model providers such as OpenAI, Anthropic, Google Gemini, Mistral, and OpenRouter. Users can get started quickly with the hosted TraceRoot Cloud service, or self-host the full platform locally via Docker, with a simple setup process that launches the interface on localhost:3000. Core platform code is released under the Apache 2.0 license, with select enterprise directories covered by a separate enterprise license. The platform supports bring-your-own-key access for all major model providers with no vendor lock-in, and connects confirmed production failures to offline evaluation loops so that every fix is validated against real-world edge cases, creating a repeatable, measurable improvement process for agent systems over successive releases.