PROVIDER - Open source, self-hostable LLM observability

Langfuse

Langfuse delivers open source LLM observability self-hosted with 1 Docker Compose stack that keeps data in-house, modelling telemetry in 3 layers of trace, session and score with prompt versioning and OpenTelemetry compatibility across 5 or more frameworks; self-hosting upkeep and thin Chinese material are the gaps.

AI Platform Germany 2026.11.19
4.3pts

Detailed Notes

Pros 5 items

Open source at the core, self-hostable with 1 Docker Compose stack that keeps data in your environment
Models telemetry in 3 layers - trace, session and score - reconstructing multi-turn sessions and human labels
Includes prompt management with versions, labels and A/B comparison, deployable from 1 change
Compatible with OpenTelemetry and mainstream SDKs across 5 or more frameworks
Provides cost and latency dashboards split across 2 dimensions, model and user

Cons 4 items

Self-hosting requires maintaining databases and object storage
Cloud and self-hosted editions release features at different cadences
Query performance at large trace volumes needs tuning
Chinese documentation and community content are relatively thin

Service Overview

Langfuse is an open source LLM engineering platform focused on tracing, prompt management and evaluation, offered as cloud and self-hosted.

In the AI platform category it is the self-controllable observability layer, replacing closed SaaS where data must stay inside the network.

Key Strengths

  • Open source and self-hosted: 1 Compose stack
  • Three-layer model: trace, session, score
  • Prompt management: versions and comparison
  • Open protocols: OpenTelemetry compatible
  • Cost visibility: per-model attribution

Product Ecosystem

Platform capabilities

Trace collection, session aggregation, scoring and human annotation, prompt management and version comparison, dataset evaluation plus cost and latency analytics.

Integrations

Ships Python and JS SDKs, supports OpenTelemetry and reaches 5 or more frameworks such as LangChain, LlamaIndex and Dify.

Typical scenarios

Intranet LLM monitoring, prompt iteration management, session quality assessment and cost attribution.

Limitations

  • Ops cost: databases and storage to run
  • Edition drift: cloud and self-hosted diverge
  • Query performance: tune at scale
  • Chinese material: limited docs and community

Use Cases

  • Data must stay internal (★★★★★): self-hosting is essential
  • Prompt version management (★★★★☆): built-in and complete
  • Multi-framework stacks (★★★★☆): protocol compatibility
  • Zero-ops preference (★★★☆☆): cloud edition is simpler

Pricing

Item Price Notes
Open source self-hosted $0 Bring your own servers and storage
Cloud Hobby $0 Free tier with baseline quota
Cloud Core / Pro Monthly subscription Higher quotas and team features

FAQ

What does self-hosting require?

A database, object storage and a container runtime; the vendor publishes a Docker Compose template.

Does it support frameworks beyond LangChain?

Yes, through SDKs or OpenTelemetry across 5 or more frameworks.

Can prompts be A/B tested?

Prompt versions and labels let you compare variants in parallel.

Is data uploaded to the cloud?

In self-hosted mode data stays entirely in your environment.

Code Examples

# Self-host: clone the official Compose stack and start
git clone https://github.com/langfuse/langfuse.git
cd langfuse && docker compose up -d
from langfuse import Langfuse

langfuse = Langfuse(public_key="pk", secret_key="sk", host="http://localhost:3000")

trace = langfuse.trace(name="qa-flow")
gen = trace.generation(name="answer", model="qwen3-32b", input="How do I cut inference cost?")
gen.end(output="Batch processing plus caching saves 50%")
langfuse.flush()

Community & Resources

Visit Langfuse Website → https://langfuse.com/
Visit Website ↗