Opik

Comet maintains Opik as a self-hostable platform for watching what LLM apps and agents do. A Python SDK (pip install opik) and a TypeScript SDK send nested traces through a @track decorator, integrations for LangChain, LlamaIndex, CrewAI, Google ADK and Agno, or plain OpenTelemetry. On the server side, a Java service and a React web app backed by ClickHouse, MySQL and Redis store those traces and add datasets, side-by-side experiment runs, LLM-as-a-judge scoring for hallucination or moderation, a prompt playground with versioned prompts, online scoring rules for live traffic, and a PyTest hook for CI. Extra packages cover an agent optimizer that tunes prompts automatically, a guardrails service, and an MCP server so coding agents can query traces. Running it locally means cloning the repo and starting opik.sh or opik.ps1, which launches everything in Docker Compose; a Helm chart handles Kubernetes. The whole stack, backend included, is Apache-2.0, so commercial self-hosting needs nothing from Comet, which also sells a hosted cloud. A sensible pick for teams that want LangSmith-style visibility while keeping trace data in house.