Langfuse

What is Langfuse?

Langfuse is an open-source engineering platform for improving language-model applications and AI agents. It captures hierarchical traces across model calls, retrieval, tools, and application steps, then lets teams filter behavior by user, session, model, latency, cost, and metadata. This makes it easier to reproduce failures and understand where an agent took an unexpected path.

The platform also manages prompts separately from application code, supports versioning and rollback, runs datasets and experiments, and attaches evaluation scores from rules, LLM judges, user signals, or human reviewers. It can be used through Langfuse Cloud or self-hosted with published deployment guides.

How to Use Langfuse

1. Add the SDK, OpenTelemetry integration, framework integration, or API instrumentation.
2. Send traces that include model, retrieval, tool, session, user, and cost context.
3. Inspect failed or expensive traces and group recurring patterns.
4. Version prompts and compare candidates in the playground or an experiment.
5. Build datasets from representative production cases.
6. Add automated and human evaluations, then monitor scores and regressions.

Core Features

• LLM and agent tracing: Records nested calls, tools, retrieval, tokens, latency, and cost.
• Prompt management: Versions, releases, fetches, caches, and rolls back prompts.
• Evaluation: Supports custom scores, LLM-as-judge, user feedback, and human review.
• Datasets and experiments: Replays examples and compares changes systematically.
• Metrics and monitoring: Tracks quality, usage, cost, and operational trends.
• Broad integrations: Works with common model SDKs, frameworks, and OpenTelemetry.
• Self-hosting: Provides deployment paths for Docker, Kubernetes, and major clouds.

Use Cases

• Debug incorrect tool calls or retrieval in an AI agent.
• Compare prompt or model changes against a fixed dataset.
• Monitor production quality, latency, and spend by customer or feature.
• Collect user and reviewer feedback for future evaluation.
• Keep production prompts versioned outside the application release cycle.

Pricing

Langfuse Cloud Hobby is free with 50,000 units per month, 30 days of data access, two users, and limited platform features. Core starts at $29 per month, Pro at $199 per month, and Enterprise at $2,499 per month, with included usage and plan-specific retention and security. Self-hosting is available, but infrastructure and operations costs remain the user's responsibility.

Frequently Asked Questions

Does Langfuse run the agent itself?

Its primary role is engineering, observability, prompt management, and evaluation around an application; the application and models can run elsewhere.

Can it evaluate production traces?

Yes. Evaluations can use automated scoring, LLM judges, user feedback, and human annotations.

Is self-hosting supported?

Yes. The official documentation lists Docker, Kubernetes, and cloud-specific deployment guides.