- CollectionObservability & Tracing
- Classificationagent-native
- LicenseApache 2.0 (open-source)
- MCPAvailable
- Agent SkillAvailable
- InterfacesAgent Skill · MCP · CLI · SDK · http
- Verified2026-10-19
What it does
Kitaru records or imports production agent runs as sessions, then replays them against the agent’s real code so a team can test a model, prompt, or code change against traffic that already happened. Tool calls are answered from the recording, so replay does not touch live systems.
The live README: “Kitaru (来る, ‘to arrive’) gives you replay-based evals for AI agents.” The product page: “Your agent’s best eval data is already in production” — turn traces into replayable evals.
The loop in docs: record or import → replay → improve (investigations, evaluators, cohorts, experiments). Kitaru sits beside an observability stack (Langfuse, LangSmith, Braintrust, Logfire, Arize Phoenix remain the system of record); it re-runs traces rather than replacing dashboards.
Agent-facing interfaces
- Catalog protocols: Agent Skill, MCP, CLI, SDK, http
- MCP: ✅ Available (kitaru[mcp] extra)
- Agent Skill: ✅ Official
- Onboarding type:
agent-skill
How agents onboard
uv add "kitaru[cli,worker,mcp]" kitaru-pydantic-ai # or: pip install
kitaru login --local # or: kitaru login <your-team-url>
npx skills add zenml-io/kitaru-skills
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| Python / TypeScript SDK | Same server; adapters for PydanticAI, LangGraph, OpenAI Agents SDK, Mastra, Vercel AI SDK |
| CLI | kitaru login, kitaru session import, experiment/replay commands; JSON when a shell is the right tool |
| MCP | kitaru-mcp — bounded operations for coding assistants |
| Agent Skills | npx skills add zenml-io/kitaru-skills |
| Docs | https://docs.zenml.io/kitaru |
Verification
Last verified 2026-10-19.
Sources:
- https://github.com/zenml-io/kitaru