kitaru

Agent traces you can run, not just read.

★ 29222 forksPythonApache-2.0zenml-io

Kitaru records or imports production AI agent traces as sessions, then replays the real agent code against changes in model, prompt, or tools so you see what improved or regressed before shipping. Self-hosted, Apache 2.0, with Python/TypeScript SDKs, framework adapters, MCP server, and agent skills.

Use Cases

Replay production agent runs against a new model, prompt, or code changeImport existing traces from Langfuse, LangSmith, Braintrust, Logfire, or Arize PhoenixBuild a regression suite that gates agent changes in CIRecord live agent sessions with a single wrapper adapterCalibrate LLM evaluators from pinned human judgments on tracesCompare baseline replays vs forked replays to isolate a changeDrive the eval loop from Claude Code, Codex, Cursor, or Windsurf via MCP and skillsSelf-host an agent trace and eval server with no user code executing on it

Built With

Language: Python
Frameworks: PydanticAI · LangGraph · LangChain · Deep Agents · OpenAI Agents SDK · Claude Agent SDK · Mastra · Vercel AI SDK · FastAPI · Pydantic · Langfuse · LangSmith · Braintrust · Logfire · Arize Phoenix · Docker/Helm

kitaru

Use Cases

Built With

Tags