Langfuse
Screenshots
1 / 2The Story
Langfuse is an observability platform for applications built on large language models. It records traces of model calls - the prompt, the response, the latency, the token count, the cost - so a team can see what an application actually did rather than inferring it from logs.
The tracing model follows the shape of an LLM application. A trace is a single request; spans inside it represent retrieval, tool calls, and model invocations. Because the structure is explicit, a slow or expensive step is visible rather than buried in an aggregate.
Beyond tracing, Langfuse provides evaluation and prompt management. A dataset of inputs can be run against a prompt and scored, either by a model or by a person, so a change to a prompt has a measured effect. The platform is published as open source and can be self-hosted, which matters for teams whose prompts contain sensitive data.
Overview
Langfuse treats an LLM application as a traceable system rather than a black box.
A trace represents one request. Inside it, spans mark the steps: a retrieval call, a tool invocation, a model request. Each span records its input, output, timing, and token usage, so the cost and latency of a request are attributable to a specific step.
Evaluation builds on that record. A dataset of inputs can be run against a prompt and scored, which turns a prompt change from a guess into a measurement. Scores can come from a model or from a person reviewing a sample.
Prompt management keeps versions of a prompt alongside the traces that used them, so a change in output quality can be traced back to a change in the prompt. The platform is open source and self-hostable for teams whose prompts contain sensitive data.
Key Features
- Structured traces
Spans separate retrieval, tool calls, and model requests so cost and latency are attributable.
- Datasets and scores
A dataset can be run against a prompt and scored, turning a change into a measurement.
- Prompt versioning
Prompt versions sit alongside the traces that used them.
- Self-hostable
The open-source core runs on your own infrastructure.
Where Langfuse is a strong choice, and where it is not
Choose Langfuse if you are running an LLM feature in production and need to know which step is slow, which is expensive, and whether a prompt change helped.
Look elsewhere if you have not shipped an LLM feature yet, or if your application makes a single model call whose cost and latency you can already see.
Getting started
Tracing starts as soon as the SDK is added; getting value from it depends on what you log.
- ~15 minTime to first trace
Install the SDK, set the keys, and the first model call appears in the trace view.
- FreeCost to start
Self-hosted is free. The cloud free tier covers a small volume of traces.
- Postgres for self-hostPrerequisites
An API key for the hosted version, or Docker and Postgres to self-host.
Known limitations
Constraints confirmed from the project's documentation.
- Self-hosting means operating the stack. Postgres and the application are yours to run and upgrade.
- Value depends on instrumentation quality. A trace is only useful if the application attaches the prompt, the retrieved context, and the output.
- Not an evaluation framework on its own. It stores scores and datasets; the evaluator is something you bring.
- Trace volume grows quickly. A chatty application produces a lot of data, which affects both storage and cost.
Pick Langfuse if you need to see what was actually sent to the model
It answers the question conventional monitoring cannot: which prompt, with which retrieved context, produced this output.
The product's value is diagnostic. When an answer is wrong, the trace shows whether the retrieval returned the wrong documents or the model mishandled the right ones, which is the difference between a fixable bug and a mystery.
If you only need aggregate cost and latency, an existing APM may already cover it and Langfuse would be redundant.
FAQ
What is Langfuse?
Langfuse is an observability and evaluation platform for LLM applications, recording traces of model calls along with evaluation and prompt management.
Is Langfuse open source?
Yes. The core is published in a public repository with a license file. Check the repository for current license terms.
Can I self-host Langfuse?
Yes. The open-source core can run on your own infrastructure, which matters when prompts contain sensitive data.
Is Langfuse free?
Langfuse lists a free tier alongside usage-based paid plans. Check the pricing page for current limits.
What stack is Langfuse built with?
CompassPad records Langfuse as a TypeScript and Next.js application on PostgreSQL with ClickHouse for analytics, based on the public repository.
What we checked
- LIVESite and primary link
The marketing site resolved and the primary product link reached the product itself.
- PRICINGPricing page
The pricing page resolved and lists the current plans and limits.
- DOCSDocumentation
Public documentation resolved and covers the product's core surfaces.
- CHANGELOGChangelog page
The linked changelog page resolved and is reachable without an account.
Discussion
No comments yet. Be the first.
Join the conversation. Sign up to comment.
Sign up free