LLM and agent tracing, evals and prompt management with an MIT-licensed core you can self-host, plus a cloud tier free to 50k units a month.
Engineering teams that want OpenTelemetry-native tracing with a credible exit: Langfuse states that all core features and APIs are in the MIT-licensed OSS build without limits, and that self-hosting uses the same deployment infrastructure as Langfuse Cloud. It fits teams happy to run Postgres and ClickHouse themselves, or teams on cloud who need a specific data region, since US, EU and Japan are selectable from the Hobby tier up. Also a reasonable landing point if you already emit OTel spans from frameworks like CrewAI, Pydantic AI or Spring AI and want them parsed into traces rather than re-instrumented.
Volume buyers need to read the billing unit carefully: a unit is any tracing data point, which means traces, observations (spans, events, generations) and scores all count, so one multi-step agent run can burn dozens of units and the 50k Hobby allowance disappears fast. Retention is where the plans really split - 30 days on Hobby, 90 days on Core at $29/mo, and 3 years only from Pro at $199/mo, with configurable retention management restricted to Pro and Enterprise. Self-hosters should not assume parity: nine features require LANGFUSE_EE_LICENSE_KEY on both the web and worker containers, including audit logs, server-side data masking, data retention policies, project-level RBAC and SCIM. On cloud, SSO (Okta, EntraID), SSO enforcement and project-level RBAC need the Pro Teams add-on at $300/mo on top of Pro's $199/mo.
Pros
Cons
The core is. The GitHub README states the repository is MIT licensed except for the ee folders, and the self-host page states all core Langfuse features and APIs are available in the MIT build without limits. Nine features sit outside that and need a LANGFUSE_EE_LICENSE_KEY: project-level RBAC, protected prompt labels, data retention policies, audit logs, server-side data masking, UI customisation, organisation creators, Org Management API with SCIM, and the Instance Management API.
Langfuse bills tracing data points, not requests. A unit is any trace, observation or score, and observations include spans, events and generations. That matters for agents: a single run that fans out across tool calls and model calls produces many observations, each billable. The Hobby allowance of 50,000 units a month and the 100,000 units included on paid plans should be read with that multiplier in mind.
Data access is tier-dependent on cloud: 30 days on Hobby, 90 days on Core at $29 a month, and 3 years on both Pro at $199 a month and Enterprise at $2,499 a month. Configurable data retention management is a Pro and Enterprise feature, and when self-hosting it is one of the features behind the enterprise licence key. Self-hosted retention is otherwise limited only by your own storage.
Yes. Langfuse receives traces on the /api/public/otel OTLP endpoint, with the traces signal at /api/public/otel/v1/traces, authenticated by HTTP Basic Auth using a base64-encoded public:secret key pair. It accepts OTLP over HTTP in JSON and protobuf, and the docs note gRPC is not supported yet. You can export from any OpenTelemetry SDK, the Collector, or instrumentation libraries such as OpenLLMetry, OpenLIT, Arize and MLflow.
Facts sourced from the vendor’s own site, last verified 2026-10-10.
The Weekly Source
One email a week. New listings, repriced tools, and the ones we would skip - including this one if its pricing changes.
