Langfuse

What it is

Langfuse is an open-source platform for tracing, evaluating and managing prompts in LLM applications and agents. You instrument the app, every run lands as a tree of observations (generations, spans, tool calls), and you score those runs with LLM judges, code evaluators or human annotation queues. Prompts live in Langfuse with versions and deployment labels, so a prompt change is something you can roll back. The core is MIT-licensed and self-hostable without usage limits. The site claims 90B+ observations a month and 21 of the Fortune 50 as users.

ClickHouse acquired Langfuse on 2026-01-16. The founders’ post says the team joined ClickHouse, the roadmap stays the same, Cloud keeps “same endpoints, same contracts”, and there are “no planned changes to licensing”. Langfuse v4, which rebuilt the data model around observations for faster queries at scale, went live on Cloud and GA for self-hosting in August 2026.

The ops question it answers: your team runs internal agents (a lead-research agent in RevOps, a contract-intake assistant in legal, a support triage bot) and nobody can see what they did, what they cost per run, or whether last week’s prompt edit made them worse.

What you are actually buying

Tracing. SDKs for Python and JS/TS plus OpenTelemetry ingestion, so it is framework-agnostic. Native integrations cover the common agent frameworks and builders such as Dify. v4 added full-text search across traces, a new filter bar and alerts.

Evals. LLM-as-a-judge (including categorical scores since March 2026), code evaluators, annotation queues for human review, and versioned datasets for experiments. Evaluators can grade images and audio since August 2026.

Prompt management. Versioned prompts, deployment labels, bulk JSON import and export between projects.

MCP server and CLI. Every plan includes an authenticated MCP server at /api/public/mcp with read and write tools, so Claude Code or another agent can list prompts, query observations and pull experiment results. CLI 1.0 shipped 2026-08-28.

Pricing

Published at langfuse.com/pricing, checked 2026-11-04. A unit is any data point you send: a trace, an observation, or a score.

• Hobby — $0. 50k units a month with no overage option, 30 days of data access, 2 users.
• Core — $29/month. 100k units included, unlimited users, 90 days of data access.
• Pro — $199/month. 100k units included, 3 years of data access, SOC 2 and ISO 27001 reports, BAA available for HIPAA. The Teams add-on is $300/month for enterprise SSO, SSO enforcement and project-level RBAC.
• Enterprise — $2,499/month. Pro plus Teams, audit logs, SCIM, uptime and support SLAs.

Overage on every paid plan is graduated: $8 per 100k units up to 1M, $7 to 10M, $6.50 to 50M, $6 above. Langfuse’s own examples: 200k units on Core is $37/month, 5M on Pro is $551/month. Data regions are US, EU and Japan. Early-stage startups get 50% off the first year.

Self-hosted OSS is free with unlimited units and users, and includes enterprise SSO. Self-hosted Enterprise has no list price: it is bundled with ClickHouse Cloud, BYOC or Private, and priced “additive to your ClickHouse commercial plan”.

The real band: a 12-person team that wants SSO lands on Pro plus Teams at $499/month before overage, against $468/month in seats alone for 12 people on LangSmith Plus and $249/month flat on Braintrust Pro. Without SSO, the same team pays $29/month on Core. Seats never move the Langfuse bill; units do.

Best for

A platform or ops engineering team running agents on more than one framework, where data residency, volume or a security review makes a metered SaaS the wrong default, and someone on the team can operate Postgres, ClickHouse, Redis and S3. Also the cheapest serious hosted option for a small team: Core at $29 with unlimited users is enough for a first production agent.

Not for

A team with nobody to run infrastructure that also needs project-level RBAC and audit logs: on self-host those need the Enterprise licence key, and on Cloud they start at $499/month. Not for a team that has not shipped an agent yet. And not if the evaluation itself is the product you want automated: Braintrust’s Loop and LangSmith’s Engine go further on generating datasets and fixes from failures.

Versus the alternatives

• LangSmith — pick it when you build on LangChain or LangGraph and want deployment and the Fleet agent workspace from the same vendor, with fewer than about ten people needing seats.
• Braintrust — the fastest-growing rival ($800M valuation in February 2026). Pick it when evals, not tracing, are the job and reviewers outnumber engineers; every plan has unlimited users.
• Arize Phoenix — the other open-source option. Pick it when you already run Arize for classical ML monitoring.
• Datadog LLM Observability — pick it when the company already pays for Datadog and the security team will not approve another vendor.
• For voice agents, Coval tests what none of these simulate.

Does the ClickHouse deal change the open-source bet? Not the licence. The signal to watch is commercial: self-hosted Enterprise is now sold on top of a ClickHouse plan. The MIT core is safe to adopt; budget for paid governance features arriving through ClickHouse’s sales motion.

Watch-outs

• Units count observations, not runs. One agent run with 15 spans and 3 scores is 19 units, so the 100k included on Core can be about 5,000 runs. Guard: run one week of real traffic on Core, divide units by runs to get your multiplier, and set the SDK sample rate (LANGFUSE_SAMPLE_RATE) on high-volume, low-risk agents before you scale.
• Self-hosting means running ClickHouse. The stack is two containers plus Postgres, Redis or Valkey, ClickHouse and blob storage, and v4 changed the data model underneath. Guard: deploy with the Helm chart or Terraform modules, pin versions, read the migration notes before each upgrade, and name an owner who is on call for it.
• Free self-host is not the compliance setup. Project-level RBAC, data retention policies, audit logs, server-side data masking and SCIM need the Enterprise licence key. Guard: list the controls your security review requires before choosing OSS, and price Enterprise against Cloud Pro plus Teams.
• The MCP server writes by default. It authenticates with a project-scoped API key and exposes write tools, so an agent with that key can change prompts. Guard: give agents keys for a dedicated project, use protected deployment labels (Teams add-on or Enterprise) for production prompts, and treat the key like a deploy credential.