Langfuse

Now live: Langfuse x Hermes integration. Hermes Agent by Nous Research is a self-improving AI agent with persistent memory, autonomous skill creation, and support for any LLM provider. It ships with a bundled Langfuse observability plugin that automatically captures every conversation turn, LLM request, and tool call, including token usage and cost. All details in the comments.

Now live: Langfuse x OpenAI Codex tracing integration. Full observability for your Codex sessions: every prompt, tool call, shell command, and token tracked. Also available for Claude Code. Link in comments.

Quarterly Langfuse Town Hall. Join the Langfuse team, maintainers, and community for our quarterly call. We'll talk about Langfuse v4, ClickHouse, cover the latest releases, roadmap, and do Q&A. 𝗪𝗵𝗲𝗻: Thursday, June 11th, 2026, at 9am PT / 6pm CEST 𝗪𝗵𝗲𝗿𝗲: Virtual on Google Meet Link in comments.

Meet us at AI Engineer Melbourne!

Last week, Marc & Leonard traveled to Munich to attend the first CDTM Job Fair organized by the CDTM Venture Fund. Lots of good conversations about Langfuse, "How to Build Good AI Agents", and why working for startup in AI is the best thing you can do right now. + We're hiring, link in comments.

𝗗𝗮𝘆 𝟱 𝗼𝗳 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗲𝗲𝗸: Langfuse MCP. Until this week, the hosted Langfuse MCP server only exposed prompt management. Today, it covers most of Langfuse: observations, metrics, scores, score configs, datasets, and their items and runs, comments, annotation queues, models, media, and health. Agents such as Claude Cowork or Linear can now investigate a production issue, pull the relevant observation, query metrics, drop a comment for the team, create a score, or create dataset items for regression testing, all without leaving the chat. The MCP server complements the Langfuse Skill and the Langfuse CLI. Use the CLI when your agent can run bash and pre-filter data. Use the MCP server when it cannot. Restrict to read-only by allow-listing lookup tools if you don't want writes. Shoutout to Ben Bachem for shipping this. Link in comments.

𝗗𝗮𝘆 𝟰 𝗼𝗳 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗲𝗲𝗸: Code evaluators. Not every evaluation needs an LLM. JSON parseability, schema validation, exact match, required tool arguments, custom business rules: things you can verify with code, you should verify with code. Deterministic, reproducible, no token cost. You can now write an 𝘦𝘷𝘢𝘭𝘶𝘢𝘵𝘦 function in Python or TypeScript directly in Langfuse, attach it to live observations or a dataset experiment, and the result lands as a native Langfuse score. Code evaluators sit alongside LLM-as-a-Judge rather than replacing it. Code wins for objective checks. A judge wins for quality, tone, or to add reasoning. They complement each other and give you a better, more complete picture together. Link in comments.

𝗗𝗮𝘆 𝟯 𝗼𝗳 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗲𝗲𝗸 𝟱: Full-Text Search When something breaks in your AI app, you need to pull the one trace that says "refund failed" out of hundreds of GB of production data. Scroll-and-hope doesn't cut it at that scale. Full-text search is now live on Langfuse Cloud. In our benchmarks, large input/output searches that took 18 seconds and scanned 494 GB now return in under half a second and read less than a gigabyte. Built on top of ClickHouse's new full-text search and a tight feedback loop with their team, helping us land features in Langfuse weeks after they ship in ClickHouse core. Humans benefit in the UI. Agents benefit in the API: the new 'matches' operator on Observations API v2 lets coding agents and scripts run token-based search programmatically. Link in comments.

𝗗𝗮𝘆 𝟮 𝗼𝗳 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗲𝗲𝗸 𝟱: the Langfuse agent skill. Building an agent is the easy part. Getting it to production is hard. You set up tracing and evaluators, but how do you know what your agent's real failure modes are? How do you know your LLM-as-a-judge is actually capturing what you care about? The Langfuse Skill turns Langfuse into a headless platform you can drive in natural language. It follows the Open Agent Skills standard, so it works with Claude Code, Cursor, Codex, and anything else that speaks the format. In the video below, Marlies Mayerhofer uses the LLM-as-a-Judge calibration skill with Codex to produce a full analysis with accuracy, F1, precision, recall, and cost, all graphed directly in the new Langfuse Experiments view.

𝗗𝗮𝘆 𝟭 𝗼𝗳 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗲𝗲𝗸 𝟱: Experiments in CI/CD. Every change of prompt, model or chunking is a potential regression you may only catch in production when it's too late. The new Langfuse Experiment GitHub Action runs your experiments on every pull request, tests your app against a dataset you control, fails the workflow when scores drop below your threshold, and posts pass/fail directly to the PR. Every run is tracked in Langfuse, so when something does regress, the failing items are only one click away. Link in comments.