Langfuse is an open-source (MIT core) observability and evaluation platform for LLM applications and agents, covering tracing, LLM-as-judge and human evaluation, prompt management, and datasets and experiments, self-hostable or consumed as Langfuse Cloud, and part of ClickHouse since January 2026. Facts below verified as of 2026-09-16.
Langfuse is the MIT-licensed observability and evaluation column the category lacked, and its open-core boundary is drawn in the filesystem: MIT outside the ee/ directories, proprietary inside them.
What it is #
A self-hostable web platform (Docker Compose locally, Kubernetes templates for production) or Langfuse Cloud, instrumented through Python and JS SDKs, OpenTelemetry, an LLM gateway such as LiteLLM, or the docs’ 100-plus library and framework integrations. Traces, sessions, users, and costs land in a UI where agents render as graphs; evaluation attaches to the same data through managed LLM-as-judge evaluators, custom code evaluators, user feedback, manual annotation queues, and custom score APIs. Prompt management versions and deploys prompts with server and client caching, a playground iterates on them, and datasets and experiments run pre-deployment evaluations against them. Made by Langfuse (YC W23), acquired by ClickHouse in January 2026, with the root LICENSE copyright now reading ClickHouse, Inc.
Status #
Mature and busy: 34,675 stars, 3,776 forks, created 2023-05-18, pushed 2026-09-16, v4.36.1 released 2026-09-15 as of 2026-09-16. Announced ClickHouse acquisition landed on Hacker News on 2026-01-17 with 220 points, and the README says the team doubled in the six months before it. Three years old, the largest community in this category, and now owned by a database vendor whose observability stack it slots into.
Strengths #
- The most permissive core license in the observability tier: MIT lets you self-host, modify, and even sell what ELv2-licensed rivals reserve.
- The full loop in one platform: tracing, evaluation, prompt management, datasets, and experiments, natively integrated rather than bolted together.
- OpenTelemetry-based ingestion keeps exit costs visible, and an agent skill, CLI, and MCP server let coding agents work with it directly.
- Real production pull: the pricing page names Canva, Twilio, Ramp, and Khan Academy, and the docs ship an interactive demo project.
Cautions #
- The
ee/split is structural: SSO, RBAC depth, and other enterprise features live in proprietary modules, so “MIT” describes the core, not the edges. - Ownership changed hands, and the acquisition thread carries the standard open-core worries (“fine as long as all these open source projects stay open source”) alongside consolidation takes about data vendors absorbing observability.
- Cloud billing counts every trace, observation, and score as a unit, then charges $8 per 100k beyond the plan allowance, so high-volume tracing bills creep.
- It observes applications you instrument, not coding sessions a harness already wrote, so it does not answer this blog’s session-analytics questions.
Pricing #
Self-hosted: free, MIT core, with the ee/ directories under their separate license.
Langfuse Cloud as of 2026-09-16: Hobby free (50k units/month, 30-day retention), Core $29/month, Pro $199/month (Teams add-on $300/month), Enterprise $2,499/month, plus graduated overage from $8 down to $6 per 100k units.
Compared to #
- Phoenix: the closest feature peer; choose Langfuse when the MIT core matters and you want prompt management built in, Phoenix for OpenInference instrumentation breadth and the Arize AX upgrade path, remembering Phoenix’s core is ELv2.
- deepeval: a pytest-style framework that gates CI on eval failures; choose it when the eval suite must block merges, Langfuse when production traces need evaluating and the team lives in a dashboard.
- LangSmith: the commercial incumbent; a Launch HN commenter described rejecting it over coupling concerns with the vendor’s framework, which is the independence argument Langfuse runs on.
Bottom line #
Recommended for teams that want production tracing, evaluation, and prompt management in one self-hostable platform under a permissive license, accepting open-core edges and ClickHouse ownership. Not for pytest-style CI gating (use deepeval) or for observing coding-agent sessions a harness already wrote on disk (that is the session-analytics category’s job).
Changes #
- 2026-09-16 - Created.
See also #
- Phoenix - the closest peer and the license contrast that decides between them
- deepeval - the eval-framework alternative when the gate is CI
- Evaluation and Review Feature Matrix - the category comparison this note joins
- Hamel Husain - the eval-driven methodology these platforms implement
References #
https://github.com/langfuse/langfuse - repository, description, stars and forks, pushed date as of 2026-09-16
https://raw.githubusercontent.com/langfuse/langfuse/main/LICENSE - MIT core, the
ee/carve-out, and the ClickHouse, Inc. copyrighthttps://langfuse.com/docs - platform overview, OTel basis, evaluation methods, agent skill, CLI, and MCP server
https://news.ycombinator.com/item?id=46656552 - the acquisition thread, 220 points, with the open-core and consolidation concerns