Langfuse

Langfuse sits between your application and the model providers as an observability and evaluation layer. Because it is built on OpenTelemetry, traces can arrive from its native Python and JS/TS SDKs, from any OTel-instrumented language (Go, Java, .NET, Ruby, PHP, Swift are named), or through an LLM gateway such as LiteLLM. Its integrations page lists native support for a long list of agent frameworks - LangChain/LangGraph, Vercel AI SDK, Google ADK, Pydantic AI, OpenAI Agents, Agno, CrewAI, DSPy, Mastra, Semantic Kernel, Strands, Temporal and more - plus direct model-provider integrations, gateways, no-code builders (Dify, Flowise, Langflow, n8n), and tracing for coding agents including Claude Code, Cursor, Codex and GitHub Copilot. The platform is open source and can be self-hosted, which is the main structural difference from the closed observability vendors it competes with.

• Hierarchical tracing of every LLM call, tool invocation and retrieval step, with session and user tracking
• Agent graphs - visualising agentic workflows as a flow rather than a flat log
• Prompt management: versioning, label-based deployment and rollback without code changes
• Datasets and experiments, runnable via SDK, UI or CI/CD
• Human annotation queues for building golden datasets
• Custom dashboards, monitors and alerts on cost, latency and quality
• A documentation MCP server and a published agent skill for AI coding assistants

Engineering teams shipping LLM or agent features to production who need to debug non-deterministic behaviour, and who want the evaluation loop (datasets, experiments, judges, human review) in the same place as the traces. The self-host option makes it a natural fit for teams with data-residency or procurement constraints that rule out a hosted-only vendor.

This is a developer tool, not a GTM tool - you instrument code, and a non-engineering revenue team has nothing to point it at. The free Hobby plan is genuinely limited for production use: 30 days of data access, 2 users, 1 annotation queue, 2 monitors, and a 30 requests/min general API rate limit, so any real team is on Core ($29/mo) at minimum and more realistically Pro ($199/mo) once it needs longer history and workable rate limits. The security controls most enterprises treat as non-negotiable are stacked high: Enterprise SSO, SSO enforcement and project-level RBAC arrive only via the Pro plan's $300/mo Teams add-on or the $2,499/mo Enterprise tier, and audit logs and the SCIM API are Enterprise-only. Self-hosting avoids the licence cost but moves the operational burden onto you.

Setup assessment: 30m

Cloud signup is free with no credit card, and for a supported framework instrumentation is an SDK import plus keys - first traces within the hour. Self-hosting, or retrofitting good trace structure across an existing codebase, is a materially larger project.

Prerequisites

• An application you can add instrumentation to (Python or JS/TS native SDK, or any OTel-capable language)
• A Langfuse Cloud account, or infrastructure to self-host
• A choice of data region if you are on Cloud - US, EU or JP, with a HIPAA-ready region on Pro and above

• A machine-readable OpenAPI specification - none of the conventional paths resolved
• The definition of a billable "unit" (the docs list a Billable Units page we did not retrieve)
• Self-hosted pricing tiers - the pricing page has a self-hosted tab whose contents were not in our excerpt
• The Langfuse skills repository contents and star count
• MCP tool names for the product MCP server
• A machine-readable OpenAPI specification was not included in the reviewed source set.