Langfuse review: Open-source tracing, evals and prompt management for LLM apps and AI agents

What Langfuse does

Records traces of LLM and agent runs, then lets you evaluate outputs, manage prompts and track cost and latency in dashboards.

Langfuse is an open-source platform that brings tracing, prompt management, evaluation and monitoring for LLM applications and AI agents into one place. Its traces record each model call, tool invocation and retrieval step, and can be filtered by user, session, cost, latency or custom metadata. Teams can keep prompts outside their code, roll them out or back, try them in a playground against real inputs, and compare models side by side.

For quality work, Langfuse supports LLM-as-a-judge evaluators, heuristic functions and human review, along with datasets, experiments and annotation queues. Dashboards and alerts track cost, latency and quality. It accepts data from any language that supports OpenTelemetry, has native Python and TypeScript SDKs, and lists more than 100 integrations. You can use the hosted Langfuse Cloud or self-host it, since the product features are MIT licensed.

Pricing on Langfuse Cloud is in US dollars. Hobby is free, with 50k units per month, 30 days of data access and 2 users. Core costs $29 per month and includes 100k units, 90 days of data access and unlimited users. Pro costs $199 per month and adds 3 years of data access and higher rate limits. Enterprise costs $2499 per month and includes the Pro features plus the Teams add-on. Usage beyond the included 100k units is $8 per 100k units on paid plans, with lower rates at volume. Discounts are listed for startups, students, non-profits and open-source projects. The pricing page does not show separate yearly-billed prices.

• Hierarchical tracing of LLM calls and tools
• LLM-as-a-judge and human evaluation
• Prompt management with rollbacks
• Playground for comparing models
• Experiments and datasets
• Dashboards and alerts
• Human annotation queues
• Self-hosting under MIT license

Our take

Langfuse fits teams building LLM apps or agents that want tracing, prompt versioning and evaluation in one tool, and that care about open source or self-hosting under an MIT license. The free Hobby plan needs no credit card, so a small project can start without paying. The trade-offs are in the limits: Hobby has 2 users, 30 days of data access and modest rate limits, and the paid plans charge for usage above the included units. Some team features such as enterprise SSO and a private support channel on Pro need the $300/mo Teams add-on. Teams that only need general application monitoring, with no interest in LLM-specific evaluation, may find much of it unused. We have not tested the product; this summary comes from the official pages.