Langfuse

What it is

Langfuse is an open source platform for LLM engineering: observability (traces, spans, generations), evaluation, prompt management, a playground for iterating on prompts, and dataset management for testing. It sits in the same layer as an APM tool, but built around the primitives that matter for LLM apps — prompts, completions, token usage, latency, cost, and eval scores — instead of HTTP requests and DB queries.

It's SDK-based (Python, JS/TS, plus OpenTelemetry-compatible ingestion) and integrates with the usual suspects: LangChain, LlamaIndex, OpenAI SDK, Vercel AI SDK. You self-host it or use their managed cloud.

Who builds it and why

Built by Langfuse GmbH, a company formed around this product — not a side project of a bigger org donating code for goodwill. The GitHub activity backs that up: 181 contributors, commits as recent as July 2026, and a release cadence in the hundreds (v3.203.1) that indicates a team shipping continuously, not a dormant repo with a README.

The "why" is straightforward: teams building LLM products hit a wall fast without tracing and eval infra, and the incumbents (Datadog, generic APM) don't model prompts/generations/evals as first-class concepts. Langfuse built for that gap specifically, open source first, commercial cloud/enterprise tier on top.

Production readiness signal

Strong GitHub signal: 30.2k stars, active contributor base (181), commits landing within the current week of any check. That's a live, well-maintained project, not abandonware.

Two things to flag before you commit:

• License is NOASSERTION. GitHub couldn't detect a clear SPDX license identifier in the repo. This isn't necessarily a red flag — Langfuse has historically used a mixed model (some components Apache-2.0/MIT, some under a BSL-style license for enterprise features) — but you need to actually read the LICENSE files in the repo before you self-host, especially if you're building a commercial product on top of it. Don't assume "open source" means "no restrictions."
• CNCF maturity: not publicly available — because it isn't a CNCF project. If you were expecting a CNCF sandbox/incubating/graduated designation as a governance signal, that doesn't apply here. Governance is company-controlled, not foundation-controlled.

Neither of these is disqualifying, but they change your risk calculus versus a CNCF-graduated project with clear Apache-2.0 licensing.

Who should use this

• Teams shipping LLM-powered features (chatbots, agents, RAG pipelines) who need to see what prompts actually went out, what came back, and what it cost — in production, not just in a notebook.
• Teams doing prompt iteration at any real scale, where "copy-paste into a Google Doc" prompt management has broken down.
• Teams that need evals (LLM-as-judge, human annotation, custom scoring) wired directly to production traces instead of a disconnected eval script.
• Anyone who wants to self-host for data residency/compliance reasons and is comfortable running a TypeScript/Postgres/ClickHouse stack.

Who should NOT use this

• Teams already deep in Datadog/Honeycomb/generic OTel tooling who just need basic LLM call logging — adding a second observability stack for token counts alone is overhead you don't need; check if your existing APM has an LLM plugin first.
• Teams uncomfortable with the licensing ambiguity — if your legal/procurement team requires a clean, unambiguous OSI-approved license before approval, get written clarification from Langfuse first; don't assume.
• Very small projects / single-developer prototypes — the self-hosted stack (multiple services, ClickHouse, Redis, Postgres) is real infra to operate. If you're pre-product-market-fit, their cloud free tier is the pragmatic choice, not self-hosting.
• Teams wanting a fully managed, zero-ops solution with enterprise SLAs baked into open source — you'll end up on their paid cloud anyway, so evaluate it as a commercial product, not a free one.

Alternatives

• Helicone – lighter-weight, proxy-based LLM observability, simpler setup, narrower eval/prompt-management scope.
• LangSmith – LangChain's own observability/eval platform; tighter integration if you're already all-in on LangChain, but closed/commercial-first.
• Arize Phoenix – open source LLM observability with strong eval tooling, more ML-observability lineage than app-engineering lineage.

Pricing

Not fully open source in the simple sense — core platform is source-available with a mixed license (verify exact terms per component; GitHub reports NOASSERTION, so don't take "open source" at face value). Self-hosting is free of licensing fees for the OSS-licensed components. Managed cloud has a free tier plus paid tiers for higher volume, SSO, and enterprise support — check langfuse.com for current pricing, it changes.