Langfuse
by Langfuse (acquired by ClickHouse, Inc., January 2026)
Open-source observability, evals and prompt management for LLM apps and agents
Langfuse is an open-source LLM engineering platform that traces every call, tool invocation and retrieval step in an AI application, then layers evaluations, prompt versioning and datasets on top. It is aimed at engineering teams running agents in production who want their trace data under their own control, and it can be self-hosted for free under an MIT licence or bought as managed cloud.
Langfuse is an open-source LLM engineering platform for tracing, evaluating and improving AI applications and agents. It was founded by Marc Klingen, Max Deichmann and Clemens Rawert and built out during Y Combinator's W23 batch, starting as open-source product analytics for LLM apps and growing into the default observability layer for teams that want trace data under their own control. The core model is hierarchical tracing: every LLM call, tool invocation and retrieval step in a request is captured as a nested span, filterable by user, session, cost, latency or arbitrary metadata, so a failing agent run can be replayed step by step rather than guessed at from logs. Around that sit prompt management with versioning and one-click rollback, evaluations run as LLM-as-a-judge, heuristic functions or human annotation queues, datasets and experiments for pre-deploy regression testing, and cost and latency dashboards with alerting. The v3 architecture splits storage across Postgres for transactional data, ClickHouse for analytics, Redis for queues and S3-compatible object storage for event payloads — the reason it scales, and the reason self-hosting is heavier than a single-binary tool. The repository carries roughly 34,000 GitHub stars under an MIT licence (excluding `ee` folders), with Python and TypeScript SDKs, native OpenTelemetry compatibility and 40+ framework integrations including LangChain, LlamaIndex, LiteLLM, CrewAI, Haystack and the Vercel AI SDK. ClickHouse acquired Langfuse on 16 January 2026 alongside a $400M Series D, committing to keep it MIT-licensed and self-hostable; ClickHouse reports 19 of the Fortune 50 and 63 of the Fortune 500 among users, naming Intuit, Twilio, 7-Eleven and Merck.
Platform or AI engineering leads shipping LLM agents to production who need per-step traces and evals, and whose security or cost position rules out sending every prompt to a closed SaaS.
A replayable, span-level record of every agent run — with cost, latency and eval scores attached — that you can host yourself for free under MIT.
At a Glance
- Category
- Developer Tools
- Pricing
- Freemium, Usage-based, Subscription, Free
- Target Market
- Platform Engineers, AI/ML Engineers, Enterprise Developers, CTOs, Data Scientists
- Deployment
- Open-source, Self-hosted, Cloud-first, Multi-cloud
- Founded
- 2023
- Headquarters
- Berlin, Germany
- Team Size
- 11-50
- Customers
- 2,000+ paying customers; 19 of the Fortune 50 and 63 of the Fortune 500 reported as users
Key Features
- ✓Hierarchical tracing
Captures every LLM call, tool invocation and retrieval step as a nested span, so a failed agent run can be inspected step by step instead of inferred from application logs.
- ✓Prompt management
Versions prompts outside application code with one-click deploy and rollback plus client-side caching, letting non-engineers iterate without a redeploy.
- ✓Evaluations
Runs LLM-as-a-judge, heuristic functions, user feedback or human annotation queues against production traces, turning quality from a vibe check into a tracked metric.
- ✓Datasets and experiments
Builds test sets from real traces and replays candidate prompts or models against them, so regressions are caught before a change ships.
- ✓Cost and latency analytics
Breaks spend and p95 latency down by user, session, model or feature, which is what makes an LLM bill actionable rather than a single opaque line item.
- ✓Free self-hosting under MIT
Docker Compose, Helm and Terraform templates for AWS, GCP and Azure with no paywalled core, so trace data never has to leave your own infrastructure.
- ✓OpenTelemetry-native integrations
Native OTel support plus 40+ framework SDKs including LangChain, LlamaIndex, LiteLLM and CrewAI, so instrumentation is usually a few lines rather than a rewrite.
Capabilities
Use Cases
- •Root-causing an agent failure
An engineer opens the trace for a failed run and sees which tool call returned malformed data three steps before the user-visible error.
- •Gating a prompt change on evals
A team replays a candidate prompt against a 200-case dataset in CI and blocks the deploy when the judge score drops below the current baseline.
- •Controlling LLM spend
A platform team attributes runaway token cost to a single retry loop in one feature and fixes it, rather than negotiating a bigger model budget.
- •Self-hosted observability under compliance constraints
A regulated organisation deploys Langfuse on its own Kubernetes cluster so prompts containing customer data never leave its network boundary.
- •Human review of production quality
Support and domain experts annotate sampled traces through annotation queues, and those labels become the dataset that future evaluations run against.
Ideal For
Best For
- ✓Debugging multi-step agent runs by replaying the exact tool calls, retrievals and model responses that produced a bad output
- ✓Regression-testing prompt changes against a saved dataset before they reach production
- ✓Attributing LLM spend and latency to specific users, sessions or features rather than one aggregate API bill
- ✓Running LLM-as-a-judge or human-annotation evaluations as a standing quality gate on production traffic
- ✓Teams in regulated or data-sensitive environments that must keep prompt and completion data inside their own infrastructure
Not Ideal For
- ✗Teams that cannot operate ClickHouse — the v3 self-host path requires Postgres, ClickHouse, Redis and S3-compatible storage, and that single constraint rules out self-hosting entirely
- ✗Small projects wanting a single-binary drop-in; independent comparisons describe the four-service architecture as meaningfully heavier than lighter-weight observability tools
- ✗Organisations that want observability bundled into an existing APM contract rather than a separate AI-specific platform
- ✗Buyers who need SCIM, audit logs and a support SLA on a small budget — those sit in the $2,499/month Enterprise cloud tier
Integrations
Deployment
Market & Ratings
2,000+ paying customers; 19 of the Fortune 50 and 63 of the Fortune 500 reported as users
Market Analysis
Pros
- ✓Genuinely open source — MIT licence with no paywalled core, and ClickHouse publicly committed to keeping it that way post-acquisition
- ✓Very large practitioner base: ~34k GitHub stars, 23.1M SDK installs a month and 6M Docker pulls, so integration problems tend to already have answers
- ✓Framework-agnostic with native OpenTelemetry support and 40+ integrations, which avoids locking observability to one agent framework
- ✓Dramatically cheaper than LangSmith at scale, and the self-host path removes per-trace pricing entirely
Cons
- ✗Self-hosting is not lightweight — v3 requires Postgres, ClickHouse, Redis and S3-compatible storage; independent write-ups say inability to run ClickHouse rules out the self-host path outright
- ✗The v2-to-v3 migration is where self-hosters get stuck, with GitHub threads reporting memory errors and timeouts
- ✗The true self-host cost is hidden: a few hundred dollars a month of infrastructure plus a fraction of an engineer owning upgrades, scaling and backups, none of which appears on the pricing page
- ✗Enterprise governance features (audit logs, SCIM, PrivateLink, support SLA) jump straight to $2,499/mo with nothing in between Pro and Enterprise
- ✗Acquisition risk is now a live consideration — the licence commitment is public, but roadmap priority for a product inside a $15B data-infrastructure company is not guaranteed
Pricing
Self-hosted (open source)
$0
- ✓MIT licence, core features free
- ✓Docker Compose, Helm, Terraform for AWS/GCP/Azure
- ✓You operate Postgres, ClickHouse, Redis and S3-compatible storage
Hobby (cloud)
$0
- ✓50k units/month included
- ✓30 days data access
- ✓2 users
- ✓No credit card required
Core (cloud)
From $29/mo
- ✓100k units/month included, then $8 per 100k
- ✓90 days data access
- ✓Unlimited users
- ✓In-app support
Pro (cloud)
From $199/mo
- ✓100k units/month included
- ✓3 years data access
- ✓Unlimited annotation queues
- ✓SOC 2 and ISO 27001 reports, BAA available for HIPAA
Enterprise (cloud)
From $2,499/mo
- ✓Audit logs, SCIM API, custom rate limits
- ✓AWS PrivateLink
- ✓Uptime and support SLA
- ✓Named lead support engineer
Cloud pricing is a monthly platform fee plus graduated usage: every paid tier includes 100k units/month, then $8 per 100k up to 1M, $7 to 10M, $6.50 to 50M and $6 above that. Core is $29/mo, Pro $199/mo with SOC 2 and ISO 27001 reports and an optional $300/mo Teams add-on, Enterprise $2,499/mo. Governance features — audit logs, SCIM, AWS PrivateLink, custom rate limits and a support SLA — are Enterprise-only, so a compliance requirement moves the price by two orders of magnitude. Self-hosting is genuinely free under MIT, but independent comparisons put real infrastructure at a few hundred dollars a month at meaningful trace volume plus a fraction of an engineer to own upgrades and backups.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 4 on domains other than langfuse.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Raindrop
AI agent monitoring that catches silent production failures: Sentry for AI agents
GitLab Duo Agent Platform
Agentic AI across the whole GitLab DevSecOps lifecycle: planning, coding, code review, CI/CD and security agents under one governance model
CodeRabbit
AI code review and agentic change management for teams shipping human- and machine-written code
Qodo
Agentic code review and governance layer for teams shipping AI-generated code at scale