Langfuse

Langfuse is an open-source AI engineering platform designed to help teams build, observe, evaluate, and improve LLM and AI-agent applications. It provides a unified environment for tracing application behavior, managing prompts, evaluating model outputs, running experiments, monitoring production performance, and analyzing usage and costs.

Langfuse was publicly launched in July 2023 as an open-source project focused initially on observability for LLM applications. The platform has since expanded into a broader AI engineering suite covering the development lifecycle from experimentation through production monitoring.

The platform’s core capability is LLM and agent observability. Langfuse captures traces that show the sequence of events occurring inside an AI application, including LLM calls, tool invocations, retrieval operations, embeddings, API calls, and other application steps. This is particularly useful for applications with complex agentic workflows where understanding how an output was produced can be difficult.

Langfuse can represent agents as graphs and track multi-step conversations as sessions. Developers can inspect individual traces and timelines to investigate latency, errors, unexpected model behavior, and other issues. The platform also supports user-level tracking, allowing teams to analyze usage and costs across individual users or applications.

Another major component is prompt management. Teams can create, edit, version, test, and deploy prompts through Langfuse. Prompt versions can be linked to application traces, making it possible to understand how a particular prompt performs in production.

The LLM Playground allows developers to test prompts and models interactively. Teams can compare different prompt configurations and evaluate their behavior before deploying changes to production. Langfuse also supports experiments against datasets so that prompt and model changes can be evaluated systematically.

Evaluation is another central part of the platform. Langfuse supports multiple evaluation approaches, including LLM-as-a-judge, code-based evaluators, human feedback, manual labeling, and custom evaluation pipelines. This allows teams to measure application quality using criteria that are appropriate for their particular use case.

Teams can create datasets containing representative examples and use them to test prompts, models, or application versions. Experiments can then compare results across different configurations, helping developers identify which changes improve quality.

Langfuse also supports evaluations on production traces. This makes it possible to monitor the quality of an application after deployment rather than relying exclusively on pre-production testing.

The platform provides human annotation workflows through annotation queues. Human reviewers can assess AI outputs and provide structured scores or labels. These human judgments can complement automated evaluations and provide additional information when quality cannot be reliably measured by an automated evaluator alone.

Langfuse also includes analytics and dashboards. Teams can monitor metrics such as latency, usage, cost, evaluation scores, and application performance. Alerts can be configured to notify teams when selected metrics cross defined thresholds. Alerts can be sent through channels such as Slack, GitHub Actions, or webhooks.

Cost and usage tracking is another important capability. Langfuse tracks the usage and cost associated with LLM calls and can break this information down by model, user, use case, or other tags. It includes pricing definitions for popular models from providers such as OpenAI, Anthropic, and Google, while also allowing teams to add custom model definitions.

This cost visibility can help development teams understand how much different AI workflows cost and identify expensive operations. The Metrics API can also expose aggregated usage and cost information for analytics, billing, and rate-limiting workflows.

Langfuse is built around an API-first architecture. Its public API exposes Langfuse functionality for custom applications and integrations. Developers can manage traces, evaluations, prompts, datasets, projects, and other resources programmatically.

The platform provides native SDKs for Python and JavaScript/TypeScript. Developers can also send telemetry using OpenTelemetry and connect applications through Langfuse’s broad ecosystem of integrations.

Langfuse supports more than 100 integrations with libraries and frameworks, making it possible to add observability to existing AI applications without completely redesigning their architecture. It also supports popular AI development frameworks such as OpenAI SDK, LangChain, and LlamaIndex.

The platform is open source and self-hostable. Langfuse’s self-hosted open-source edition is available under the MIT license and includes the core platform features, APIs, observability, evaluation, prompt management, and datasets. Organizations can therefore run Langfuse on their own infrastructure rather than using the hosted cloud service.

Self-hosting can be particularly useful for organizations with strict data-management requirements or teams that want greater control over infrastructure and deployment. Enterprise self-hosted options add capabilities such as project-level RBAC, data-retention policies, audit logs, and enterprise support.

Langfuse Cloud provides the managed alternative. It removes the infrastructure-management burden while providing the same general platform capabilities through a hosted service.

The current cloud platform offers several pricing tiers. The Hobby plan is free and includes 50,000 units per month, 30 days of data access, and two users. The Core plan costs $29 per month and includes 100,000 units per month, 90 days of data access, and unlimited users. The Pro plan costs $199 per month and provides three years of data access, higher rate limits, and additional enterprise-oriented capabilities.

Langfuse also provides an Enterprise plan at $2,499 per month according to the current public pricing page. Enterprise adds features such as audit logs, SCIM API, custom rate limits, uptime and support SLAs, and a dedicated support engineer.

The cloud pricing system is usage-based in addition to the plan fee. Langfuse defines a billable unit as a trace, observation, or score. The total number of units is calculated as traces plus observations plus scores.

For the Core, Pro, and Enterprise tiers, additional usage currently starts at $8 per 100,000 units, with lower rates at higher volumes. This means organizations should consider both the subscription tier and expected application telemetry when estimating total costs.

Langfuse’s approach is intentionally not seat-based. The company states that usage-based pricing better reflects the value of the platform because AI engineering is inherently collaborative. Its open-source edition and free Hobby cloud plan also provide entry points for small projects and experimentation.

Langfuse is particularly useful for AI engineering teams, developers, ML engineers, and organizations operating production LLM applications. Its capabilities become increasingly valuable as AI systems become more complex and require systematic monitoring, testing, evaluation, and cost control.

For AI agents, Langfuse can capture traces across multiple steps, including tool calls, retrieval, model generations, and other operations. This gives developers a way to understand how an agent reached a particular result and where problems occurred.

The platform also supports RAG applications. Retrieval operations, embeddings, and LLM calls can be included in traces, allowing developers to investigate retrieval quality and understand the relationship between retrieved information and generated responses.

Langfuse’s combination of observability, prompt management, evaluation, datasets, experiments, and analytics creates a continuous development loop. Developers can deploy an application, observe production behavior, identify problems, create tests, experiment with changes, evaluate results, and then deploy improvements.

The platform’s open-source nature is another important differentiator. Teams can use the hosted cloud service for convenience or self-host the platform when they need greater infrastructure control. The company also maintains a public API and supports custom integrations.

Langfuse became part of ClickHouse in January 2026, with the company stating that the combination would accelerate its development of the AI engineering platform.

Overall, Langfuse is best suited to teams that need more than basic LLM logging. Its combination of observability, evaluation, prompt management, experiments, cost tracking, datasets, and open-source deployment makes it a comprehensive platform for developing and operating production-grade AI applications.

Pricing

Langfuse offers both cloud-based paid plans and a free open-source self-hosted edition. The current cloud plans are Hobby, Core, Pro, and Enterprise. Hobby is free, Core starts at $29/month, Pro is $199/month, and Enterprise is $2,499/month.

Cloud usage is additionally measured through billable units consisting of traces, observations, and scores. Additional usage is charged according to the applicable tier.

Pricing Model

Paid

Pricing Note

Langfuse has a free Hobby plan, paid cloud plans starting at $29/month, and a free self-hosted open-source edition. Cloud pricing is based on plan subscription plus usage measured in traces, observations, and scores.

Review

Langfuse is one of the more comprehensive platforms for teams developing and operating LLM and AI-agent applications. Its biggest strength is that observability, evaluation, prompt management, datasets, experiments, and cost tracking are integrated into one platform.

The tracing system is particularly useful for complex AI agents and RAG applications because it can show LLM calls, retrieval operations, tool calls, and other application steps. This gives developers much more visibility than basic application logging.

Its evaluation capabilities are another major advantage. Teams can combine LLM-as-a-judge, code evaluators, human feedback, and custom scoring rather than relying on one evaluation method.

The open-source and self-hosted option is also a significant benefit. Teams can run the core platform themselves under the MIT license, while organizations that prefer managed infrastructure can use Langfuse Cloud.

The main limitation is that Langfuse is primarily an AI engineering platform, not a general-purpose AI assistant. Getting the most from it requires developers to understand telemetry, evaluations, datasets, prompts, and LLM application architecture.

Cloud costs can also increase with application volume because pricing includes usage-based billing for traces, observations, and scores. Teams with very high telemetry volumes may need to optimize what they ingest.

For organizations building production LLM applications and AI agents, however, Langfuse provides a strong combination of visibility, testing, evaluation, and optimization tools.