Langfuse is an open source LLM observability platform that traces every call your AI agents make in production. This guide covers what it does, what running it costs and what you need to add around it: real-time guardrails, data and memory management, regulatory governance, including for your agentic data science workflows.
Langfuse: the open source LLM engineering platform
Trace, analyse and iterate on your LLM applications
Langfuse is an open source LLM engineering platform that lets teams monitor LLMs in production. It observes and captures traces of every API call to the model: latency, inference cost, tokens consumed per response and errors returned. These metrics let you debug issues and monitor performance before they affect your end users. Prompt management, automatic evaluation and evaluation data management round out Langfuse. These evaluations often rely on an LLM judge, replayed on every deployment: we measured its cost and accuracy against a decision model in Jev vs LLM-as-a-judge. With more than 80 native OpenTelemetry integrations (langfuse.com/docs, 2026), Langfuse covers the full range of modern LLM applications and engineering needs, from machine learning models to LangGraph agents with their prompt templates in production.
Langfuse, OpenTelemetry and agent frameworks
OpenTelemetry is becoming the de facto standard for observability of agentic systems in production, and Langfuse supports it natively. As an open source LLM engineering platform, Langfuse connects to every major framework: LangGraph, Google ADK, LangChain, the OpenAI SDK and LiteLLM. This convergence on an open standard means your existing Python stack gets instrumented without a rewrite. Langfuse helps teams monitor their AI agents and large language model applications, however diverse their technical ecosystem.
Langfuse joins ClickHouse: what changes for your stack
ClickHouse acquired Langfuse on 16 January 2026. The MIT licence is maintained and self-hosting remains possible in the short term. What the acquisition changes: Langfuse's cloud roadmap and observability priorities are now aligned with those of a company whose business model differs from yours. The question of long-term strategic dependency deserves an answer before any structural investment in your LLMOps stack.
Langfuse in production: what you need to add around it
Langfuse observes. It blocks nothing, does not store your agents' state and does not manage access. In production, these layers are added alongside it.
Guardrails: PII, prompt injection, topic restrictions
Langfuse traces what your agent does. Guardrails block what it should not do: PII detection, prompt injection defence and topic restrictions, the three most common risk vectors in production LLM deployments. Ideally they are set per agent, per model and per use case. The full audit trail stays available through the Langfuse API and your dashboard, in real time: Langfuse observes every token generated by your model.
Memory and session persistence: SQLite in dev, PostgreSQL in prod
Managing conversational state is the number one problem for multi-turn LLM agents. It is handled in the framework: with LangGraph, a SQLite checkpointer for local development and PostgreSQL in production. Langfuse traces and observes every user session, without storing its state on the agent's behalf. The two layers are complementary.
Framework-agnostic: LangGraph, Google ADK and the rest
Langfuse builds on OpenTelemetry: LangGraph, Google ADK, LangChain, the OpenAI SDK and LiteLLM connect to it. Apply the same logic to the rest of the stack: open standards (OpenTelemetry, MCP tools, AG-UI) rather than a tool tied to a single framework, LLM or cloud provider. Open source LLM engineering has to stay open: if the ecosystem converges on a new standard in 18 months, you do not start from scratch.
Use cases: LangGraph agents traced with Langfuse
Sales qualification agent: trace every decision, control cost per call
A LangGraph agent that pre-qualifies B2B leads generates dozens of LLM calls per conversation. Without traces, the cost latency of each exchange stays a black box. Langfuse tracing makes every model call visible: latency, cost, tool triggered and decision taken. Topic restrictions keep the agent within its sales scope, and session memory keeps context between qualification steps. The result: redundant calls identified, evaluation of quality metrics, a user feedback loop and a measurable reduction in inference cost from the first week in production.
Customer support agent: real-time guardrails, automatic PII detection
An agentic support agent with access to sensitive customer data exposes your organisation to GDPR risk on every call. Langfuse records every response generated by your LLM agent to safeguard its quality. A guardrails layer has to block PII before it reaches the model. Off-scope topics are refused with no extra application code to maintain. On the European timeline, the Article 50 transparency obligations have applied since 2 August 2026, and compliance for Annex III high-risk systems has been postponed to 2 December 2027 (Regulation (EU) 2024/1689, as amended by the digital omnibus regulation in force since 27 July 2026): the Langfuse audit trail, combined with guardrails, is a first operational answer.
N8N → LangGraph migration: observability from day one
If your agentic workflows run on N8N, Make or Zapier and move to LangGraph agents, switch Langfuse on from go-live: no black box during the transition, no blind spot after deployment. Teams that want to build and develop LLM applications from their existing automations get full visibility over their workflows from day one. Plan observability into the migration itself, not as a separate Langfuse integration phase.
What Langfuse covers, and what you need to add
Langfuse covers observability for LLM applications, and does it very well. What it does not cover: real-time data protection, conversational state management, agent governance and AI Act regulatory compliance. The table below structures this evaluation for teams in the technical selection phase.
| Criterion | Langfuse alone | To add in production |
|---|---|---|
| LLM observability (traces, metrics, evaluations) | Native | Nothing more |
| Real-time guardrails (PII, prompt injection, restrictions) | None | A dedicated guardrails layer |
| Memory and session management (conversation persistence) | To be built | Persistence in the framework (SQLite, PostgreSQL) |
| Agent governance (access, tools, sessions) | None | Access control (SSO, RBAC) and per-agent rules |
| AI Act compliance (OWASP Agentic AI, audit trails) | None | Retained traces, human oversight, reporting |
| Sovereign deployment (on-premise, cloud, hybrid) | Complex (5 components) | A chosen hosting location: on-premise, cloud or hybrid |
| Setup (infrastructure required) | PG + ClickHouse + Redis + S3 + K8s | Depends on the components chosen |
Sources: Glassbrain.dev 2026, Checkthat.ai 2026, langfuse.com/docs 2026.
The "Setup" row sums up the operational stakes: PostgreSQL + ClickHouse + Redis + S3 + Kubernetes to run when self-hosting (Glassbrain.dev, 2026). The Langfuse cloud Hobby tier covers 50,000 units/month for 2 users with 30 days of data retention (Checkthat.ai, 2026): not enough once your data engineering team grows. Langfuse observes and guarantees the quality of traces; protecting and orchestrating your workflows in production is the job of other components.
Real cost and dependency
The true cost of self-hosting Langfuse v3
After the ClickHouse acquisition: keeping reversibility
Since ClickHouse acquired Langfuse in January 2026, the question of strategic dependency comes up in every LLMOps project. The answer comes down to one principle: choose tools you stay in control of. An open licence, public and auditable code, sovereign deployment, and standard instrumentation (OpenTelemetry) that lets you switch observability backend without re-instrumenting your agents, including if the Langfuse roadmap shifts under ClickHouse's influence.
Beyond traces: measuring real usage
Langfuse tells you what each agent does. It does not tell you who uses it, for what, or what it returns next to the licences and assistants you already pay for. That is the reading Idun Group provides: the real usage of the AI already deployed, internal agents included, cross-checked with interviews, to decide what to industrialise, consolidate or cut.
