Le Répertoire des AgentsFR · EN · ES · JA · HI

“I don't know what my agents are really doing”

Developers and teams running agents in production: when an agent answers nonsense or loops for an hour, you want to replay the scene instead of guessing.

Recommended agent: Langfuse

💰 $0 to $29/month to get started (free self-hosted or with the Hobby plan; Core from $29/month). Level: Advanced 🌍 Germany Free Hobby plan (50,000 units/month) and free self-hosted open-source version; paid plans from $29/month

Price verified on 2026-10-29.

Langfuse (langfuse.com), the open-source observability platform for AI agents and applications: every run becomes a “trace” you can replay step by step — prompts, responses, tools called, durations, errors and cost in dollars. You can add automatic evaluations to spot drifts before your users do.

How it works

Trigger: One of your agents runs a task — records every step of the run, computes its cost and flags anomalies and errors, and you get a complete, searchable history of what your agents did, and what it cost.

Step-by-step setup

Create a Langfuse Cloud account (free Hobby plan) or self-host the open-source version.
Add the Langfuse SDK to your agent, or plug it in via your proxy (it integrates with LiteLLM in particular).
Open your first trace and replay a full run, step by step.
Create a cost-per-agent dashboard and an alert on abnormally long or expensive runs.

⚠️ Guardrails: Traces contain your prompts and your data: choose the hosting region (EU available), mask personal data before sending, and set a short retention period to start.

🔧 Technical details

Installation: hybrid — managed Langfuse Cloud and self-hosting on your own infrastructure, access via SDK/API in both cases.
Prerequisites: Cloud: an account, a project and a pair of API keys (public/secret), then a Python or JS/TS SDK or an OpenTelemetry exporter. Self-hosted: Docker, PostgreSQL, ClickHouse, Redis/Valkey (documented minimum version 7) and S3/blob object storage.
Hardware: Documented minimums for self-hosting: Web 2 CPU/4 GiB, Worker 2 CPU/4 GiB, PostgreSQL 2 CPU/4 GiB, Redis/Valkey 1 CPU/1.5 GiB, ClickHouse 2 CPU/8 GiB, MinIO/SeaweedFS blob 2 CPU/4 GiB or serverless S3.
Self-hosting: yes — Docker Compose for local/test use on a VM, Kubernetes via Helm (the recommended path in production) and Terraform templates for AWS, Azure and GCP.
License: mixed — all code outside the ee/ directories is under the MIT license (the repo's LICENSE file: “MIT Expat”); the modules under ee/, web/src/ee/, worker/src/ee/ and packages/shared/src/server/ee/ fall under the license defined in ee/LICENSE (commercial Enterprise license).
API & integrations: API: Yes · SDK: Python, JavaScript/TypeScript · CLI: Yes · Webhooks: Yes · MCP: Yes
Models: Model-agnostic platform: tracing of 100+ frameworks and providers (OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Azure OpenAI, LiteLLM, LangChain, LlamaIndex, etc.) and local models via the Ollama integration. For the Playground, experiments and LLM-as-a-judge, the documented LLM connections are OpenAI, Azure OpenAI, Anthropic, Google Vertex and Amazon Bedrock (other services via the OpenAI API), with the user's own API keys (BYOK).
Data: In the Cloud, data is hosted on AWS and ClickHouse Cloud in the chosen region (EU Ireland eu-west-1, US Oregon us-west-2, Japan Tokyo ap-northeast-1, plus a HIPAA region in Oregon): traces/observations/scores in ClickHouse, transactional data in Postgres, events first persisted in S3. The official security FAQ states that traces and prompts are used neither for model training nor for external benchmarks; self-hosted, data stays in the user's infrastructure, including in offline/air-gapped mode.
Documented limits: Cloud: Hobby 50k units/month, 30-day data access, 2 users and 1,000 req/min ingestion; Core 100k units included, 90 days, 4,000 req/min ingestion; Pro 3 years of access, 20,000 req/min ingestion and higher API limits (the pricing page details the rate limits per API). The self-hosted OSS is documented with no usage limit.
Operations: Native traces/observations and metrics; Google, AzureAD and GitHub sign-in, enterprise SSO (Okta, Entra ID), RBAC at organization and project level, SCIM and audit logs depending on the offers/Enterprise license. Exports via API, batch export via the interface and scheduled export to blob storage; documented Cloud backups: Postgres 7 days with PITR (RPO 10 min, RTO 12 h), ClickHouse every 6 h (4 days), S3 versioning 7 days — Docker Compose, on the other hand, is documented without high availability or backups.

Official sources: langfuse.com · github.com · github.com · langfuse.com · langfuse.com · langfuse.com — verified on November 1, 2026.

Find your agent with the quiz →

Frequently asked questions

How much does Langfuse cost?

$0 to $29/month to start (free self-hosted or with the Hobby plan; Core from $29/month). Free Hobby plan (50,000 units/month) and free self-hosted open source version; paid plans from $29/month. Price verified on October 29, 2026.

Is Langfuse free?

Yes. Free Hobby plan (50,000 units/month) and free self-hosted open source version; paid plans from $29/month.

Is Langfuse hard to get started with?

Level: Advanced. Setup takes 4 steps; the first: Create a Langfuse Cloud account (free Hobby plan) or self-host the open source version.

What limits and guardrails should I know about with Langfuse?

Traces contain your prompts and your data: choose the hosting region (EU available), mask personal data before sending, and set a short retention period at first.