Langfuse Open-Source LLM Engineering Platform

Langfuse was started in 2022 by Marc Klingen, Max Deichmann and Clemens Rawert, went through Y Combinator W23 and runs engineering from Berlin with a go-to-market office in San Francisco. It joined ClickHouse in January 2026 while keeping its product, MIT license and cloud service. It helps teams building LLM applications and agents see what their systems do in production and improve quality systematically.

The platform captures hierarchical traces of every LLM call, tool invocation and retrieval step, with filtering by user, session, cost and latency. It adds evaluations via LLM-as-a-judge, heuristics or human annotation queues, prompt management with versioned deployments and rollbacks, a playground, dataset experiments, and cost and latency dashboards with alerts. It integrates through OpenTelemetry and Python and JavaScript SDKs, and runs as Langfuse Cloud in US, EU and Japan regions or self-hosted on Docker or Kubernetes. More than 50,000 companies use it, including 21 of the Fortune 50.

• Tracing: hierarchical traces of LLM calls, tool use and retrieval steps.
• Evaluations: LLM-as-a-judge, heuristic scoring and human annotation queues.
• Prompt management: versioned prompts with one-click deployment and rollback.
• Experiments and playground: compare models and prompts on datasets side by side.
• Cost and latency analytics: dashboards and automated alerts.
• Flexible deployment: cloud in US, EU and Japan regions or self-hosted via Docker, Helm and Terraform.

Vietnamese banks can self-host Langfuse to monitor customer service chatbots and audit answers. Software companies building AI agents can run evaluations before each release. E-commerce platforms can track cost and latency of product search assistants across models.

All features are MIT licensed with no gates on self-hosting, so companies can keep trace data inside their own infrastructure. Tracing and evaluations make it faster to find hallucinations, slow steps and cost spikes in production LLM apps. Prompt management lets product teams change prompts without code releases. A free Hobby plan includes 50,000 observations per month, and the Core plan costs 29 USD per month.

As companies move LLM applications and agents into production, observability and evaluation have become essential to control quality, cost and risk. Open-source, self-hostable tooling is favored where data residency rules apply, including in Southeast Asian banking, telecom and public sectors.

Started in 2022; Y Combinator W23; 4 million USD seed in September 2023 led by Lightspeed, La Famiglia and Y Combinator. Joined ClickHouse on January 16, 2026, keeping the MIT license and cloud service. Offices in Berlin and San Francisco. SOC 2 Type II, ISO 27001 and GDPR compliant, with a HIPAA-ready region.