Langfuse

Three things get bolted together badly in most first agent deployments: where the traces go, where the prompts live and how anyone knows whether a change actually helped. All three sit in one MIT-licensed codebase here. Instrument with the Python or TypeScript SDK, or point an existing OpenTelemetry exporter at the endpoint, and every trace, observation and score lands in ClickHouse where the dashboards query it. Prompts get versions and labels and are fetched through a client-side cache, so a prompt lookup does not add latency to the hot path.

Evaluation runs in both directions. Offline, you build a dataset and run experiments against it to catch regressions before a prompt change ships. Online, LLM-as-judge scorers and human annotation queues attach scores to real production traces, which is usually the only way to discover that retrieval quality quietly degraded when someone rebuilt the index. Scores are first-class data rather than a bolt-on, so cost, latency and quality appear on the same timeline. The v4 release of 17 August 2026 added full-text search across observation inputs and outputs, deterministic Python and TypeScript checks on live observations, and threshold alerts that fire into Slack, webhooks or GitHub Actions.

Licensing deserves precision. As of June 2025 every product feature is MIT, including the ones competitors normally gate behind a paid tier. Only SCIM, extended audit logging and data retention policies live in clearly marked /ee directories that need a commercial key, and the core works without them. Cloud starts free at 50k units a month and tops out at $2,499 for enterprise. ClickHouse acquired the company on 16 January 2026 as part of a $400M Series D that tripled its valuation to $15B, with a public commitment that the roadmap, the licence and self-hosting all continue unchanged. Worth watching. Not yet worth worrying about.