Langfuse

Researched 2026-06-28. Primary category: llm-observability.

One-liner — Open-source LLM engineering platform for tracing, evaluating, and managing prompts of AI apps; now owned by ClickHouse.

What it does — Langfuse instruments an LLM/agent application and records every step (calls, tool use, latency, token cost) as traces. On top of that it offers evals (LLM-as-a-judge + custom scores), prompt management/versioning, a playground, and datasets for regression testing. It is the open-source default many teams reach for to answer “what did my agent actually do, what did it cost, and is output quality regressing?” Self-host the whole stack or use Langfuse Cloud.

Where it sits in the stack — The llm-observability layer of the AI model/prompt tier. This is a development- and operations-time visibility tool, not an inline prompt or egress control, though traces help you detect prompt-injection or data-leak incidents after the fact. Lives alongside (and downstream of) the ai-gateway and ai-runtime-security controls.

Deployment & architecture — SDK/OpenTelemetry-based instrumentation (Python, JS/TS) plus integrations with LangChain, the OpenAI SDK, litellm, and others. Backed by ClickHouse for trace storage (the reason for the acquisition — “LLM observability is fundamentally a data problem”). Fully self-hostable (MIT-licensed core) or SaaS. Integrates into CI for eval gating.

Positioning & differentiators — Strongest open-source community position in the category, OpenTelemetry-native, broad framework-agnostic instrumentation. Versus langsmith (tied to the LangChain ecosystem, closed-source) Langfuse is framework-neutral and self-hostable; versus arize-phoenix it overlaps heavily on OSS tracing/evals; versus helicone it is SDK/OTel-based rather than proxy-based; versus braintrust it is open-source and observability-first rather than eval-first.

Ownership, funding & M&A — Acquired by ClickHouse, Inc., announced 2026-01-26 (confirmed via deal-counsel announcement; terms undisclosed). Langfuse continues as an open-source product inside ClickHouse, pitched as part of a combined open-source build/monitor/optimize stack. Founded 2022 (YC W23) by Max Deichmann, Clemens Rawert, Marc Klingen; HQ Berlin with SF presence. Pre-acquisition funding modest (~$4.5M reported; backers Lightspeed, La Famiglia, General Catalyst, YC). Confidence: high on the acquisition fact and date.

CTO / hedge-fund lens — Day-1 if you are building or buying any internal LLM/agent app and want cost + quality visibility without a Datadog-scale contract. The self-host option is attractive for a fund that wants traces (which may contain prompts/MNPI-adjacent content) to stay inside its own boundary. ClickHouse ownership lowers vendor-continuity risk versus a tiny startup. Not a model-risk/SR 11-7 governance tool on its own — pair with an ai-governance-platform for that.

Competitors / alternatives — langsmith, arize-phoenix, braintrust, helicone, datadog, comet.

Open questions / to verify

• Exact acquisition price (undisclosed).
• Whether Langfuse Cloud pricing/SLAs change under ClickHouse.
• Reconcile founding year (2022 per founders/YC W23 vs “2023” cited in acquisition release).

Sources

• Langfuse Acquired by ClickHouse, Inc. (Orrick deal announcement) — fetched 2026-06-28 — supports: acquisition by ClickHouse, date 2026-01-26, open-source continuity, customer/funding background; confidence: high.

History

• [2026-06-28] Stub created from seed registry.
• [2026-06-28] Researched; established Langfuse acquired by ClickHouse (2026-01-26, confirmed) — changed ownership independent→acquired, confidence low→high; filled founding (2022, YC W23), Berlin HQ, OSS deployment, funding. No seed M&A flag existed but acquisition found and recorded.