Langfuse

Open-source LLM engineering / observability platform (YC W23, now part of ClickHouse as of Jan 2026, ~31k GitHub stars). Backed by ClickHouse for storage. Open-core: MIT-licensed core, with an "ee/" (enterprise edition) subtree under a separate license — self-hostable free tier is the full MIT core, not a crippled trial.

Core Features

• Tracing/observability — instrument an app to capture LLM calls, retrieval, embedding, and agent-action traces; inspect complex multi-step sessions
• Prompt management — centrally version-controlled prompts with server/client caching (no added latency)
• Evaluations — LLM-as-judge, code evaluators, user feedback capture, manual labeling, custom eval pipelines via API
• Datasets — test sets/benchmarks for continuous eval, pre-deployment testing
• LLM Playground — jump from a bad trace straight into iterating on the prompt/model config
• Full API + typed Python/JS SDKs, OpenAPI spec

Deployment

• Managed: cloud.langfuse.com (generous free tier)
• Self-host: docker compose up (local, 5 min), VM, Kubernetes/Helm (preferred prod), Terraform modules for AWS/Azure/GCP

Integrations

OpenAI SDK (drop-in), LangChain, LlamaIndex, Haystack, LiteLLM, Vercel AI SDK, Mastra, Amazon Bedrock, Ollama, AutoGen, CrewAI, smolagents, Instructor, DSPy, and more.

• GitHub: https://github.com/langfuse/langfuse
• Site: https://langfuse.com/

Related

• Loom — Loom's built-in OTel tracing/cost-tracking covers similar ground for AgentCore-based agents; Langfuse is the framework-agnostic, dedicated-platform equivalent
• Headroom — complementary layer: Headroom compresses context going into the LLM, Langfuse observes what actually happened around the call
• LiteLLM
• Strands_Agent
• MCP
• Agent_Harness