Langfuse
🪢 Open source LLM engineering platform: LLM Observability, metrics, evals, prompt management, playground, datasets. Integrates with OpenTelemetry, Langchain, OpenAI SDK, LiteLLM, and more. 🍊YC W23
Last verified:
What is Langfuse?
Langfuse is an open-source LLM engineering platform that helps teams collaboratively debug, analyze, and iterate on their LLM applications. It provides comprehensive observability through hierarchical traces that capture every LLM call, tool invocation, and retrieval step, along with prompt management, evaluation tools, playground, datasets, and analytics dashboards all in one integrated platform.
Key features include LLM application and agent tracing with traces and graphs, session tracking for chats/threads, user tracking, token and cost tracking, native SDKs for Python and JavaScript, OpenTelemetry support, and 80+ integrations with frameworks like LangChain, OpenAI SDK, LiteLLM, and LlamaIndex. The platform also offers prompt versioning, prompt fetching, prompt release management with labels, prompt composability, server and client-side prompt caching, an interactive playground, and prompt experiments. For evaluation, Langfuse provides datasets, experiments via SDK and UI, custom evaluation scores, user feedback tracking, LLM-as-judge evaluators, and human annotation queues.
Langfuse is built for developers, AI engineers, ML teams, and organizations building AI applications and agents from prototype to production. It serves over 100,000 engineers and 2,300+ customers including 19 Fortune 50 companies, processing 10+ billion observations per month. The platform is particularly valuable for teams who value open-source, want control over their data, need to monitor cost/latency/quality metrics, and require enterprise-grade security and compliance including SOC 2 Type II, ISO 27001, and HIPAA eligibility.
The platform uses a ClickHouse OLAP database backend for fast querying of millions of traces in milliseconds, async ingestion via Redis queue for non-blocking tracing, S3/Blob storage for large payloads, and edge-cached prompts. Langfuse is MIT-licensed, self-hostable via Docker Compose, Kubernetes (Helm), or Terraform on AWS/GCP/Azure, and features an extensive public REST API with query SDKs and S3 blob storage exports for data portability.
Langfuse pricing
Pricing model: Freemium
Langfuse offers four pricing tiers. The Hobby plan is free with 50k units/month included, 30 days data access, 2 users, community support via GitHub, and no credit card required. The Core plan costs $29/month with 100k units/month included (additional usage at $8/100k units with volume discounts), 90 days data access, unlimited users, and in-app support. The Pro plan costs $199/month with everything in Core plus 3 years data access, data retention management, unlimited annotation queues, high rate limits (20,000 requests/min ingestion), SOC2 & ISO27001 reports, BAA available for HIPAA, and prioritized in-app support. The Teams add-on is $300/month extra and includes Enterprise SSO (e.g., Okta), SSO enforcement, fine-grained RBAC, and support via dedicated Slack/MS Teams channel. The Enterprise plan costs $2,499/month and includes everything in Pro + Teams plus audit logs, SCIM API, custom rate limits, uptime SLA, support SLA, dedicated support engineer, with optional yearly commitment for custom volume pricing, architecture reviews, billing via AWS Marketplace or invoice. Self-hosting the MIT-licensed open source version is free at all plan levels. Discounts available: 50% off for early-stage startups (first year), up to 100% off for research/students, $199/month credits for non-profits, and $300/month credits for open-source projects (first year).
Langfuse pros
- Fully open source with MIT license for self-hosting
- Hierarchical traces capture all LLM calls, tool invocations, and retrieval steps
- Native Python and JavaScript SDKs with strong typing
- 80+ integrations including LangChain, OpenAI SDK, LiteLLM, LlamaIndex
- OpenTelemetry native for standard trace format and reduced vendor lock-in
- Prompt versioning with one-click deployments and rollbacks via labels
- Interactive playground to test prompts on real production inputs
- LLM-as-a-judge evaluators for automated quality assessment
- Human annotation queues for collaborative review workflows
- Dataset management for systematic testing in development
- Token and cost tracking with detailed usage metrics
- ClickHouse backend queries millions of traces in milliseconds
- Async ingestion ensures tracing never blocks application performance
- 99.9% uptime with 10+ billion observations processed monthly
- SOC 2 Type II, ISO 27001, GDPR compliance with HIPAA eligibility
- Multi-modal tracing support for text, images, and other modalities
- Scheduled batch export to S3 blob storage for data portability
- MCP servers and CLI for coding agent integration
- Multi-region data support (US, EU, JP) with EU data residency option
- Extensive public REST API with rate-limited access for all features
Langfuse cons
- Hobby plan limited to 2 users and 30 days data retention
- Additional usage costs $8/100k units beyond included quota
- Enterprise SSO and SSO enforcement require Teams add-on ($300/mo)
- Project-level RBAC only available with Teams add-on
- Audit Logs only included in Enterprise plan
- Response time SLO is 48h for Core and Pro (not guaranteed for Hobby)
- Self-hosted FOSS version lacks SOC2 certification
- Some advanced features like Protected Deployment Labels need Teams add-on
- Enterprise plan is expensive at $2,499/month base price
- Data retention management feature only in Pro plan and above
Frequently asked questions about Langfuse
What is Langfuse and what does it do?
Langfuse is an open-source LLM engineering platform that helps teams collaboratively debug, analyze, and iterate on their LLM applications. It provides comprehensive observability through hierarchical traces that capture every LLM call, tool invocation, and retrieval step. The platform includes prompt management with versioning and deployment, evaluation tools (LLM-as-judge, human annotation, datasets), an interactive playground, experiments, and analytics dashboards. Langfuse helps you ship AI agents and products from prototype to production by bringing observability, prompts, evals, experiments, and human annotation into one connected workflow.
Is Langfuse free to use?
Yes, Langfuse has a free Hobby plan that includes 50k units/month, 30 days data access, 2 users, and all platform features with limits. No credit card is required to sign up. Additionally, the open-source MIT-licensed version can be self-hosted for free with unlimited use. The free plan is great for hobby projects, POCs, and getting started. For production projects needing longer data retention and unlimited users, paid plans start at $29/month (Core).
How do I self-host Langfuse?
Langfuse can be self-hosted easily via Docker Compose, Kubernetes (Helm chart), or Terraform on AWS/GCP/Azure. The core Langfuse is MIT-licensed and requires no license key for self-hosting. You can deploy with a simple 'docker compose up' command or apply the Helm chart/Terraform from GitHub. The self-hosted version scales to billions of monthly events and includes all core features. For enterprises needing additional features like UI customization, there is a paid Enterprise Edition (EE) available.
What integrations does Langfuse support?
Langfuse supports 80+ integrations including native SDKs for Python and JavaScript, OpenTelemetry for Java/Go/custom implementations, and framework integrations with LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Google ADK, CrewAI, LiveKit, LlamaIndex, AutoGen, DSPy, Dify, Langflow, OpenRouter, n8n, Spring AI, and many more. It works with any model provider including OpenAI, Anthropic, Amazon Bedrock, Azure OpenAI, Mistral AI, Google Gemini, xAI, vLLM, Groq, and Ollama. Langfuse is OTel native and works with any language and framework supporting OpenTelemetry instrumentation.
What is the difference between traces, sessions, and observations in Langfuse?
Traces are the top-level container that includes all LLM and non-LLM calls, including retrieval, embedding, API calls, and more - they represent a complete request or workflow. Sessions track multi-turn conversations or agentic workflows over time, grouping multiple traces together (useful for chats/threads). Observations are the individual steps within a trace, including generations (LLLM calls), spans (arbitrary logic), and events (single-point events). Traces provide lowest-level transparency, sessions allow tracking of user conversations, and observations capture the detailed steps within each trace.
How does Langfuse evaluation work?
Langfuse provides flexible evaluation tools including LLM-as-a-judge evaluators that can run fully managed on production or development traces, heuristic functions, human review via annotation queues, user feedback tracking, and custom evaluation scores (numeric, boolean, categorical). You can create and manage datasets for systematic testing in development, run experiments via SDK or UI to test new prompt versions, and apply evaluators to any step within your application for step-wise evaluations. Evaluation results can be plotted in Langfuse dashboards to monitor production health and measure output quality.
What is Prompt Management in Langfuse?
Prompt Management allows you to manage, version, and optimize your prompts throughout the development lifecycle without code changes. You can create prompts via UI, SDKs, or API; collaboratively version and edit prompts; deploy prompts to production or any environment via labels; test prompts interactively in the playground; and run experiments against datasets. Features include prompt versioning, unlimited prompt fetching, prompt release management, prompt composability, server and client-side prompt caching, protected deployment labels (Pro+ Teams), and the ability to link prompts with traces to understand performance in context. The prompt config feature stores structured data like model parameters, function definitions, and JSON schemas versioned together with the prompt.
What are Langfuse datasets and experiments?
Datasets are collections of test cases (input-output pairs) used for systematic testing in development to ensure your application performs reliably across different scenarios. Experiments allow you to test prompts and models on datasets directly in the UI without custom code, or via SDK. You can run experiments to systematically test your LLM application, compare results side by side across different prompt versions or models, and evaluate latency, cost, and evaluation metrics across versions. This enables you to test new prompt versions before deploying to production and validate changes against baseline performance.
How does Langfuse handle data security and compliance?
Langfuse is SOC 2 Type II and ISO 27001 compliant, with GDPR compliance and HIPAA eligibility. It offers data regions in US, EU, and Japan, with HIPAA-compliant regions available for Enterprise. Security features include sign-in with Google, AzureAD, GitHub, enterprise SSO (e.g., Okta, AzureAD/EntraID) with Teams add-on, SSO enforcement, client-side data masking, project-level RBAC (Teams add-on), organization-level RBAC, SCIM API for automated user provisioning (Enterprise), and audit logs (Enterprise). The platform uses async ingestion via Redis queue, ClickHouse OLAP database, and S3/Blob storage for reliable enterprise-scale processing with 99.9% uptime.
What is the Langfuse API and how do I use it?
Langfuse has an extensive public REST API where all platform data and features are accessible. API keys are available in Langfuse project settings, and authentication uses Basic Auth. Both the Langfuse Python SDK and JS/TS SDK provide strongly-typed wrappers around the REST API for convenience, accessible via the 'api' property on the Langfuse client instance. The API includes v2 data APIs (Observations API v2 for retrieving observation data, Metrics API v2 for aggregated analytics), dataset APIs, and query SDKs. Rate limits vary by plan: Hobby has 30 requests/min for general API routes, Core has 100 requests/min, Pro has 1,000 requests/min, and Enterprise has custom rate limits. Data can also be exported via scheduled batch export to S3 blob storage.