The archive · Product Ideas · Product decision · 2023–2026
Langfuse gives LLM apps software-style traces and versioned prompts, open source first
YC W23 team builds LLM engineering as open source — traces, evals, prompt versioning, self-hostable — growing to 34k GitHub stars and joining ClickHouse.
Langfuse
What it had to solve
LLM features were shipping as production software while debugging them was guesswork: a bad answer could come from any model call, retrieval step or prompt version, and nothing recorded the path. Langfuse's YC W23 founders had spent 15 months building tracing since an earlier Show HN, and their December 2024 Launch HN presented an open-source platform that wraps traces, prompt management, evals and datasets into one workflow.
How it works
Langfuse was built on an analogy: an LLM application is software, and software needs traces. When a model returns a bad answer there is no stack trace telling you which prompt, retrieval step or agent action produced it, and no version history telling you what changed. Langfuse's founders — a YC W23 team that had been working on tracing for 15 months since an earlier Show HN — made that visibility the product.
The platform records everything: typed Python and JavaScript SDKs, a drop-in OpenAI SDK, callbacks for LangChain and LlamaIndex, and an @observe decorator route every model call, retrieval and agent action into traces. On top of the traces sit prompt management with version control and low-latency caching, LLM-as-judge and code evaluators, datasets for pre-deployment benchmarks, and a playground you can reach directly from a bad trace. The core is MIT-licensed and runs self-hosted via Docker in minutes, while Langfuse Cloud offers a managed version.
The December 2024 Launch HN drew 215 points and 61 comments, with the founders reporting thousands of adopting teams including Khan Academy, Twilio and Samsara, and announcing v3, a scalability rewrite built around ClickHouse. By September 2026 the repository showed 34.2k stars and 3.7k forks, dozens of leading open-source AI projects from Langflow and Open WebUI to LlamaIndex list Langfuse among their integrations, and the README said the company had become part of ClickHouse in January 2026.
Why it lands
- The insight was a gap, not a feature list: AI apps were the only production software without observability, so tracing was the obvious missing layer.
- Open source came first: an MIT core that teams could self-host removed the trust objection for code that watches every prompt and call.
- Tracing alone would have been a log viewer; tying it to versioned prompts, evals and datasets made it a closed improvement loop.
- Sitting underneath the ecosystem's frameworks — OpenAI, LangChain, LlamaIndex, LiteLLM — let Langfuse become the shared layer other tools integrate with rather than a competing silo.
What it did
The Launch HN drew 215 points and 61 comments; the founders reported thousands of teams using the project, naming Khan Academy, Twilio and Samsara, alongside a v3 scalability rewrite. By September 2026 the repository showed 34.2k stars and 3.7k forks, dozens of leading open-source AI projects from Langflow and Open WebUI to LlamaIndex list Langfuse among their integrations, and the README announced the company became part of ClickHouse in January 2026.
What you can take
Before promising what an AI app can do, sell the ability to see what it did: observability, versioned prompts and evals made one open-source layer the whole ecosystem chose to integrate with.
Since then
Langfuse's v3 rewrite, announced in the same Launch HN, replaced Postgres upserts with an ingestion pipeline of S3 events, Redis queues and a ClickHouse worker, cutting dashboard latency to 400ms at the 95th percentile and making prompt retrieval effectively zero-latency. In January 2026 the README said Langfuse had become part of ClickHouse, with the team hiring across the EU; the core remained MIT-licensed outside an enterprise folder, and the project continued as the shared observability layer for a wide swath of open-source AI stacks.
Sources
- Launch HN: Langfuse (YC W23) – OSS Tracing and Workflows to Improve LLM Apps
- langfuse/langfuse: Open source LLM engineering platform (README)
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card