Langfuse is an open-source AI engineering platform built to help teams build, monitor, and improve large language model applications and agents. It combines observability, prompt management, evaluation tools, and analytics dashboards in one workspace. The platform covers the lifecycle of generative AI tools, from early prototypes to production deployments.
The platform addresses a specific problem. LLM outputs are non-deterministic and generate large amounts of verbose input and output data that traditional observability tools struggle to handle. When prompts are hardcoded into application logic, updating them requires engineering time and a full deployment cycle, which slows down iteration.
Langfuse asynchronously captures hierarchical traces of every LLM call, retrieval step, and tool invocation, so tracing does not block the host application. Prompts are managed in a dedicated UI where teams can edit, test, and deploy changes using labels. Quality is measured through LLM-as-a-judge evaluators, manual annotation, and custom code checks, with results aggregated in a ClickHouse-backed database.
Langfuse is MIT-licensed for self-hosting and built natively on OpenTelemetry. It also provides a Model Context Protocol server, a CLI, and ready-made skills so coding agents like Claude Code or Cursor can query traces and manage prompts directly. In January 2026, Langfuse GmbH (also operating as Finto Technologies Inc.) joined ClickHouse, Inc., while continuing to operate Langfuse Cloud and the open-source project under its existing MIT license.
Pricing
Langfuse offers a free Hobby plan with 50,000 units per month, 30 days of data access, and a 2-user limit, requiring no credit card. Paid tiers start with Core at $29 per month (100k units, 90-day data access, unlimited users), Pro at $199 per month (3 years of data access, SOC2 and ISO27001 reports), and Enterprise starting at $2,499 per month with custom rate limits and an uptime SLA. Usage beyond the included 100k units on paid plans is billed at $8 per 100k units, with lower rates at higher volume. An optional Teams add-on for $300 per month adds enterprise SSO and role-based access control to the Pro plan.
* Disclaimer: Please note that pricing information may not be up to date. For the most accurate and current pricing details, refer to the official website.
Key Features
-
✓
Hierarchical tracing for LLM calls, chains, and retrieval steps
-
✓
Session tracking for multi-turn conversations and threads
-
✓
Automated token and cost tracking by user and session
-
✓
Centralized prompt versioning, decoupled from application code
-
✓
LLM-as-a-judge evaluation plus manual human annotation queues
-
✓
Native OpenTelemetry support and an MCP server for coding agents
Use Cases
Collaborative Prompt Optimization
Product managers and domain experts edit prompts directly in the Langfuse UI without engineering involvement. Because prompts are decoupled from application code, updated versions can be deployed instantly using labels, while the application fetches the latest version through client-side caching.
LLM Application Debugging
Developers investigate unexpected agent behavior, hallucinations, or latency spikes in production. Hierarchical traces capture the exact prompt, model response, tool executions, and retrieval steps, letting developers isolate the specific step where an error occurred.
Automated Quality Monitoring
QA teams continuously assess the accuracy and safety of an AI application at scale. LLM-as-a-judge evaluators automatically score live production traces against defined criteria, catching regressions without relying solely on manual review.
Cost and Token Tracking
Engineering leaders and FinOps teams monitor the financial overhead of generative AI features. Token usage and cost are logged automatically and can be broken down by user, session, region, model, or prompt version.
Regression Testing Before Deployment
Engineering teams verify that new models or code changes do not degrade performance. Running CI/CD experiments against curated datasets lets teams compare code variants and block deployments if evaluation scores drop too low.
Agent-Driven Development
Developers use autonomous coding agents to build and instrument applications faster. The MCP server, CLI, and repository skills let tools like Claude Code and Cursor query traces and manage prompts through natural language.
Strengths & Weaknesses
Strengths
Langfuse is MIT-licensed, allowing free self-hosting with no vendor lock-in.
Tracing and prompt fetching add no measurable latency, due to asynchronous logging and client-side caching.
The ClickHouse backend can query millions of traces in milliseconds and process billions of events monthly.
Native OpenTelemetry compatibility and over 100 integrations support most existing tech stacks.
SOC 2 Type II, ISO 27001, GDPR, and HIPAA-eligible compliance are built into the platform.
Weaknesses
Historical data access is capped at 30 days on Hobby and 90 days on Core.
The free Hobby plan limits accounts to a maximum of 2 users.
Enterprise SSO and fine-grained RBAC require a separate $300 per month Teams add-on on Pro.
Hobby plan API rate limits, capped at 30 general requests per minute, can bottleneck bulk operations.
Who Is This For?
AI Engineers and Developers: need granular tracing and debugging tools to optimize LLM chains and agent graphs without adding user-facing latency.
Product Managers: need a centralized UI to write, version, test, and deploy prompts without waiting on engineering release cycles.
QA and Testing Teams: rely on LLM-as-a-judge pipelines, annotation queues, and dataset experiments to measure AI output quality.
Engineering Leaders and FinOps: need analytics dashboards to attribute API costs and track token consumption across user segments.
Frequently Asked Questions
What is Langfuse used for?
Langfuse is an open-source AI engineering platform for tracing, prompt management, evaluations, and analytics, used to build and improve LLM applications from prototype to production.
Does Langfuse slow down my application?
No. Tracing data is sent asynchronously through local queues, and prompts are cached client-side, so application response time is not affected.
Can I self-host Langfuse for free?
Yes. The core product is MIT-licensed and can be self-hosted using Docker Compose, Kubernetes, or Terraform without vendor lock-in.
How much does Langfuse cost?
The Hobby plan is free with 50k units per month. Paid plans start at $29 per month for Core, $199 for Pro, and $2,499 for Enterprise, with usage billed past 100k units.
What is a billable unit on Langfuse Cloud?
The source material does not define the exact composition of a billable unit. Plan limits and overage pricing are expressed in units of 100k per $8.
Is there a learning curve for new users?
Langfuse combines tracing, prompt caching, multiple evaluation methods, and datasets in one platform, which can take time to learn for teams new to LLM observability.
Does Langfuse support compliance requirements?
Pro and Enterprise plans include SOC 2 Type II and ISO 27001 reports, GDPR support with selectable data regions, and HIPAA-eligible environments.
Can prompts be managed without redeploying code?
Yes. Prompts can be written, versioned, and tested in the Langfuse UI, then deployed to production instantly using deployment labels.
What changed after the ClickHouse acquisition?
Langfuse joined ClickHouse, Inc. in January 2026. According to company statements, Langfuse Cloud continues running as before, and the core product stays MIT-licensed and self-hostable.
Does Langfuse work with coding agents like Claude Code or Cursor?
Yes. Langfuse provides a Model Context Protocol server and CLI so coding agents can query traces and manage prompts directly.
Langfuse integrates with OpenAI and Anthropic for native tracing of model calls, LangChain and LlamaIndex for chain and RAG pipeline tracing, LiteLLM as a proxy for multi-model logging, and the Vercel AI SDK for tracing Next.js applications. It exports metrics to PostHog and Mixpanel, and supports coding agents including Cursor, Claude Code, and Codex through its MCP server and CLI.