opik

Opik records what your AI app actually did. Every model call, agent step and tool call lands in a trace you can click through, and the same platform runs evals: test datasets, experiments, and AI-graded scores for things like hallucination. It comes from Comet, the ML experiment-tracking company, and the whole repo is Apache 2.0 with no enterprise folder.

Self-hosting is real work. The local install is one script, but it starts about eight containers (ClickHouse, ZooKeeper, MySQL, Redis, MinIO, a Java backend, a Python evaluator and the UI), and Comet says that Docker setup is not for production. Production means Kubernetes and ClickHouse nodes that want 24GB of RAM or more. Anonymous usage stats are on by default; one environment variable turns them off.

The cloud Free plan covers 25,000 spans a month for up to 10 people. Pro is $19 a month for up to 50 people and 100,000 spans, then $5 per extra 100,000. Solo and small teams: start on the free cloud. Teams that can't send prompts to a vendor: self-host. SSO, RBAC and compliance paperwork are Enterprise only. langfuse/langfuse and Arize-ai/phoenix are the closest alternatives, and Phoenix is source-available, not open source.

The catch: self-hosted Opik has no login at all. Comet keeps authentication in its enterprise product, so anyone who can reach the port can read every trace, including your prompts and any user data inside them. Put it behind a VPN or oauth2-proxy/oauth2-proxy first. It also ships almost daily, so plan on frequent upgrades.