How much RAM does Langfuse need?

Langfuse runs six services: the web frontend, a background worker, PostgreSQL, ClickHouse, Redis and MinIO. Minimum is 4 GB RAM; 8 GB is recommended for steady production traffic. ClickHouse is the heaviest component (~1–2 GB at load).

Is Langfuse free to self-host?

Yes. The Langfuse core (MIT licence) is completely free. The optional Langfuse Cloud SaaS (langfuse.com) has a free tier and paid plans. On ServOrbit, you only pay for the VPS — Langfuse itself costs nothing.

Which LLM providers does Langfuse support?

Langfuse is provider-agnostic. Native integrations exist for OpenAI, Anthropic, Azure OpenAI, Cohere, and any OpenAI-compatible API (Ollama, LiteLLM, Mistral, Groq…). LangChain, LlamaIndex, Dify, Haystack, and VercelAI also have first-class integrations.

Does Langfuse require a domain?

Yes. Langfuse uses NextAuth.js for authentication, which requires a publicly resolvable domain (HTTPS) to set the correct callback URL and session cookies. A subdomain on any domain you own is sufficient.

Build, host and run AI solutions.

Langfuse

Trace, evaluate and manage your LLM applications — self-hosted, production-ready.

4 GB (8 GB recommended for production) RAM 2 vCPU Port 3000 Available

Source code

Tech stack

DockerNext.jsPostgreSQL 16ClickHouse 24Redis 7MinIO

Minimum RAM4 GB (8 GB recommended for production)

Minimum CPU2 vCPU

Default port3000

Compatible OSChoice of Linux distributions

Langfuse is the leading open-source LLM observability platform. It gives you deep visibility into every AI request your application makes: full prompt and completion traces, latency breakdowns, cost tracking per model and per user, and structured evaluation scores — all in a single dashboard.

Deployed on your VPS, Langfuse replaces cloud-hosted alternatives (Langsmith, Helicone, Arize) without sending your traces or user data to third parties. A single SDK call (`@observe()` decorator or `langfuse.trace()`) instruments any Python, TypeScript, or OpenAI-compatible stack in minutes — LiteLLM, LangChain, LlamaIndex, Dify, and direct API calls all emit traces automatically.

Version 3.x (MIT core) adds a prompt management hub where engineers can A/B-test prompt versions, run evaluations with LLM-as-a-judge or human raters, and promote the winning prompt to production without a code deploy. A REST API, a web SDK and native integrations make it the operational backbone for any team shipping AI to users.

Key features

Full-stack LLM tracing: prompt, completion, latency, cost and token count per call — nested spans for agent chains

Native SDK for Python and TypeScript, plus automatic integration with LiteLLM, LangChain, LlamaIndex, and Dify

Prompt management: version-control your prompts, A/B test variants, push updates without code deploys

Evaluation framework: run LLM-as-a-judge, human annotation, or model-based scoring against any trace

Cost and usage analytics: breakdown by model, endpoint, user and session — compare providers at a glance

Dataset management: capture production traces as golden test sets for offline evaluation

Role-based access control: scoped API keys per environment (dev, staging, prod)

When to use this solution?

1

Debug LLM regressions

When a new model or prompt version degrades quality, filter traces by evaluation score, compare side-by-side with the previous version, and reproduce the failing input in one click.

2

Monitor production AI costs

Track token spend by model, user and feature — set alerts when a single endpoint exceeds your budget. Switch from GPT-4 to a cheaper model using data, not guesswork.

3

Iterate on prompts without deploys

Push a prompt update from the Langfuse UI, route a percentage of traffic to the new version, collect evaluation scores, and promote the winner — all without touching your application code.

Deploy Langfuse on your VPS

Guide optimized for ServOrbit Cloud VPS.

01

Order a ServOrbit VPS (4 GB RAM minimum)

Langfuse runs PostgreSQL, ClickHouse, Redis and MinIO alongside the web and worker services. A 4 GB plan is the practical minimum; 8 GB is recommended for production workloads.

02

One-click install

Select Langfuse in the ServOrbit marketplace and click Deploy. Docker Compose brings up all six services automatically. The web UI is available on your domain once all health checks pass (allow 60–90 seconds on first boot for ClickHouse initialisation).

03

Create your first project

Open https://your-domain.com and register (signup is enabled by default). Create a project and copy its public/secret key pair from the Settings page.

04

Instrument your application

In Python: pip install langfuse, set LANGFUSE_HOST=https://your-domain.com, LANGFUSE_PUBLIC_KEY and LANGFUSE_SECRET_KEY, then decorate your LLM calls with @observe(). In TypeScript: npm install langfuse, initialise with your host and keys. Traces appear in the dashboard within seconds.

05

Enable automatic tracing (LiteLLM)

If you already have LiteLLM in your stack: add success_callback = ["langfuse"] to your litellm_config.yaml and set the three env vars above. Every proxied LLM call is traced automatically — no application-level code change needed.

06

Run your first evaluation

Go to Traces, select a representative set, click 'Add to dataset'. Open Evaluations, create a score template (e.g. 1-5 relevance), attach it to the dataset, and run a batch evaluation with LLM-as-a-judge or your own scoring function.

Frequently asked questions

Langfuse is an open-source LLM observability and prompt management platform. It records every prompt sent to a language model and every completion received, then lets you analyse quality, cost, latency and regressions — all on infrastructure you control.