Langfuse
github.com/langfuse/langfuse- Category
- Developer Tools
- Rank
- No. 33Tools index
- Pricing
- Freemium
- Platform
- web
- Type
- APP
- Builder
- langfuse
- GitHub
- 33.2k stars
- Latest release
- v4.11.0
- Added
- Jul 2, 2026
About
Langfuse is an open-source LLM engineering platform that provides observability, prompt management, evaluation, and experimentation tools for AI applications. It helps teams trace every LLM call, monitor cost and latency, run evaluations, and continuously improve their AI products from prototype to production. It integrates with 100+ frameworks and model providers with no vendor lock-in.
What it does
Langfuse gives an AI team one workspace for following application behavior from request to result. It records model calls alongside retrieval, embedding, and agent activity, then connects failures to prompt iteration. Teams can version prompts, build evaluation datasets, apply automated or human judgments, and retry troublesome cases in a playground.
Why it's ranked here
The case for Langfuse is breadth with a coherent workflow. A bad trace can become a playground experiment, a dataset case, and an evaluated result. It also supports managed hosting, local Docker deployment, production Kubernetes deployment, typed SDKs, and a public API. That makes it credible for both adoption and customization.
What's good
Instrumentation covers more than model responses, including retrieval, embeddings, agent actions, sessions, and user feedback. Prompt caching aims to keep centralized prompt changes from adding application latency. Evaluation options span model judges, code evaluators, manual labels, and custom pipelines. OpenAPI, Postman, Python, and JavaScript or TypeScript clients support bespoke workflows.
Tradeoffs
Self-hosting carries meaningful infrastructure weight. The preferred production route is Kubernetes, while repository development expects Node 24, pnpm, Docker services, a database, and ClickHouse configuration. The broad shared server surface includes queues, object storage, authentication, evaluation execution, deletion processing, and analytics integrations. That scope increases operational and codebase complexity.
How to use it well
Use Langfuse when several engineers need a repeatable loop from captured failures to prompt changes, dataset tests, and evaluations. Start with framework instrumentation or typed SDKs, then add custom evaluation pipelines through the API. Choose managed hosting for less operations work or Kubernetes for control. Keep a separate model provider and application framework.
Technical notes+
The root package.json defines a private pnpm monorepo requiring Node 24 and orchestrated with Turbo; its scripts cover builds, tests, type checking, Docker-backed development infrastructure, database work, and releases. packages/shared/src/index.ts exposes a very broad shared surface spanning domain models, evaluations, datasets, prompts, pricing, query types, and Prisma types, while packages/shared/src/server/index.ts aggregates storage, ingestion, ClickHouse, Redis queues, authentication, notifications, evaluations, and repository services. packages/shared/scripts/seeder/cli.ts prechecks database and ClickHouse environment variables before dynamically importing the main seeder, avoiding an opaque schema failure on fresh clones. packages/in-app-agent-sandbox-runtime/src/server.ts implements an HTTP sandbox service with validated operations, a 10 MiB request limit, command timeouts, workspace path confinement, and lifecycle hooks.
Observed
- License
- MIT
- Implementation languages shown
- TypeScript and JavaScript
- Package management
- Private pnpm monorepo with Turbo orchestration and Node 24 requirement
- Client interfaces
- Python and JavaScript or TypeScript SDKs, public API, OpenAPI specification, and Postman collection
- Deployment surfaces
- Managed cloud, Docker Compose for local or virtual-machine hosting, and Helm on Kubernetes
- Infrastructure templates
- Terraform templates are documented for AWS, Azure, and GCP
- Core data infrastructure
- The repository states that Langfuse is built with ClickHouse
Read from README.md, package.json, packages/config-eslint/index.js, packages/shared/src/index.ts, packages/eslint-plugin/src/index.ts, packages/in-app-agent-sandbox-runtime/src/server.ts, packages/shared/src/domain/index.ts, packages/shared/src/errors/index.ts, packages/shared/src/server/index.ts, packages/shared/scripts/seeder/cli.ts, packages/shared/src/encryption/index.ts, packages/shared/src/in-app-agent/index.ts, packages/shared/src/tableDefinitions/index.ts, packages/shared/src/server/s3/index.ts, packages/shared/src/server/otel/index.ts.
What it can do
Trace every LLM call across an AI application
LLM application instrumented with Langfuse SDK or integrated framework → Detailed trace logs showing inputs, outputs, latency, and cost per call
Monitor cost and latency of LLM usage
Traced LLM calls and model provider data → Dashboards and metrics showing token usage, spend, and response times
Manage and version prompts
Prompt templates created or edited in the platform → Versioned, retrievable prompts deployable to LLM applications without code changes
Run evaluations on LLM outputs
LLM responses and evaluation criteria or human annotations → Scores and evaluation results measuring quality, accuracy, or custom metrics
Test and experiment with prompts and models in a playground
Prompt text, model selection, and configuration parameters → Live LLM responses for comparison and iteration
Build and manage datasets for testing and benchmarking
Collected traces or manually curated input-output examples → Reusable datasets for running regression tests and evaluations
Debug AI application behavior by inspecting traces
Recorded traces from a running AI application → Step-by-step breakdown of LLM calls, tool use, and chain execution for root cause analysis
Intel on Langfuse
Tags
Tech Stack
Media
Comments (0)
No comments yet
Indexed by a proprietary survey. Corrections welcome.