Review

Langfuse: open-source observability with a real platform tax

Langfuse is a strong choice for teams that need tracing, prompt management, and evaluations in one open-source stack, but the cloud pricing and governance layers add friction fast.

Last updated April 2026 · Pricing and features verified against official documentation

LLM observability stops sounding abstract the first time a production system behaves badly and nobody can tell whether the model, the prompt, the retrieval layer, or the tool call caused the mess. Langfuse exists for that moment. It gives teams a place to see traces, compare prompts, run evaluations, and keep the telemetry in a shape that engineers can actually use.

That is also why the product now feels larger than the category name suggests. Langfuse began as an open-source LLM engineering platform and, after ClickHouse acquired it in January 2026, it kept the same core promise: stay open, stay self-hostable, and make AI application debugging feel less improvised. The commercial wrapper around that promise has become more serious, though. Langfuse is now a platform buy, not just a tracing tool.

The best case for Langfuse is straightforward. If your team ships production LLM features and wants observability, prompt management, and evaluation in one stack, it is one of the more coherent options available. The open-source core, OpenTelemetry support, and self-hosting path make it unusually flexible for engineering teams that care about control.

The case against it is just as clear. Langfuse asks you to think like a platform owner even when the immediate job is just to inspect a few model calls. The pricing model uses units, the governance features are split across tiers and add-ons, and the product rewards teams that are already serious about AI operations. If you only need a light tracing layer, Langfuse is probably more machinery than you want.

What the Product Actually Is Now

Langfuse is best understood as an LLM engineering system rather than a single observability dashboard. The current product covers tracing, prompt management, evaluations, experiments, datasets, annotations, metrics, and API-driven workflows. It works through SDKs, OpenTelemetry, and a public API, and it can run as Langfuse Cloud or be self-hosted on Docker, Kubernetes, or cloud infrastructure.

That matters because the product has a split identity. On one side, it is a developer tool for instrumenting AI applications and debugging production behavior. On the other, it is a governance surface that now includes retention controls, RBAC, SSO, audit logs, and deployment options that make sense only once more than a few engineers depend on it. The acquisition by ClickHouse did not change that basic shape; it made the infrastructure posture more obvious.

Strengths

It puts the whole LLM loop in one place. Langfuse is not just tracing, and that is the point. Teams can inspect calls, version prompts, run experiments, collect feedback, and compare quality signals without stitching together a separate observability tool, a prompt registry, and a spreadsheet of eval results.

It fits real engineering stacks instead of asking for a rewrite. OpenTelemetry support, SDKs, and an API-first model make Langfuse easy to slot into existing systems. That is valuable for teams that already have application telemetry, custom orchestration, or a mix of frameworks and model providers.

The open-source and self-hosting story is genuine. Langfuse is not using “open source” as decorative marketing. The current site frames the core platform as MIT licensed, and the self-hosting docs cover Docker, Kubernetes, and cloud deployment paths. For teams that need control over where telemetry lives, that is a meaningful distinction.

The lower tiers are usable rather than symbolic. Hobby is free, Core starts at $29 per month, and both tiers are enough to evaluate the product seriously before procurement gets involved. That is better than the common SaaS trick of giving you a free login that cannot carry real work.

Weaknesses

Usage billing can get expensive without much drama. The cloud plans include a unit allowance, then add overage at $8 per 100k units, with lower rates only at higher volumes. That is a fair structure for infrastructure, but it means Langfuse can move from “reasonable” to “accounting problem” as soon as production traffic starts to scale.

The governance story is split across tiers and add-ons. Pro includes the stronger compliance posture, but the Teams add-on is where enterprise SSO, enforcement, fine-grained RBAC, and private Slack or Teams support live. That is workable, but it also means buyers who assume the headline plan covers everything will be surprised.

It is a platform, so it rewards platform thinking. Langfuse is strongest when the user already knows how they want to instrument, evaluate, and review their AI system. Teams that want a simpler answer engine or a lightweight notebook will find the setup overhead hard to justify. For those buyers, a narrower product like LangSmith or a workflow builder such as Dify may fit better.

Pricing

Langfuse’s pricing reads like a ladder from experimentation to procurement. Hobby is free, includes 50k units per month, 30 days of data access, and two users. Core is the obvious first paid tier at $29 per month, with 100k units, 90 days of data access, unlimited users, and in-app support. That is the value tier for small teams that are already in production and want more history without paying enterprise rates.

Pro jumps to $199 per month and is where the product starts to look like a serious business system. It keeps the 100k-unit allowance but extends data access to three years, adds data retention management, unlimited annotation queues, higher rate limits, and access to SOC 2 and ISO 27001 reports, with a BAA available for HIPAA use cases. If governance and retention matter, this is the tier that begins to earn its keep.

Enterprise then moves to $2499 per month, with yearly commitment options for custom volume pricing. It adds audit logs, SCIM, custom rate limits, uptime and support SLAs, and a dedicated support engineer. That tier is aimed at larger organisations that need vendor management, not just better software.

The real pricing trap is the Teams add-on. At $300 per month, it contains some of the features buyers usually assume are already in the base product: enterprise SSO, SSO enforcement, fine-grained RBAC, and private Slack or Teams support. Langfuse is still better value than many enterprise tools, but the governance cost shows up where buyers least want surprises.

Privacy

Langfuse’s privacy policy is better written for a business buyer than a consumer one. The policy says the company does not process sensitive personal information, does not receive information from third parties, and may process data to provide, improve, and secure the service. It also says customer data is handled as processor data under a DPA when Langfuse is acting on behalf of a customer.

The more useful operational details sit in the product docs and pricing pages. Langfuse says its cloud service supports GDPR, offers a DPA, and provides data retention, masking, and deletion controls. The higher tiers add compliance reports and BAA availability, and the platform can also be self-hosted or run in air-gapped setups if a team wants to keep telemetry entirely in its own environment.

I did not find a public statement saying customer traces are used to train models, and that is not how the policy frames the product. The real issue is access and retention, not model training: who can see traces, where they are stored, how long they remain available, and whether your team is buying enough tiering to get the controls it needs.

Who It’s Best For

• The platform team shipping LLM features in production. They need tracing, prompt management, and evaluations in one system, and they want something that can grow with the application instead of collapsing into a one-off debugging tool.
• The organisation that wants open source without giving up a managed service. Langfuse lets a team start in Cloud, move to self-hosting, or split the difference, which is useful when procurement and engineering do not want the same thing.
• The engineering group already living in telemetry and APIs. OpenTelemetry, SDKs, and the public API make Langfuse a good fit for teams that instrument software as a matter of habit.
• The company that needs retention and governance, but not a giant vendor suite. Langfuse gives you SSO, RBAC, audit logs, SCIM, and compliance artifacts without dragging you into a broader workspace platform.

Who Should Look Elsewhere

• Teams that only want a narrow tracing product should compare Langfuse with LangSmith first.
• Buyers who want to assemble AI workflows rather than observe them should look at Dify or n8n.
• Organisations that do not want to manage telemetry, retention, or evaluation workflows should avoid Langfuse entirely and buy something simpler.

Bottom Line

Langfuse is one of the strongest open-source choices for teams that need to understand what their LLM applications are actually doing in production. It covers the right surfaces, it integrates in the right ways, and it gives engineering teams enough control to be useful in real systems rather than only in demos.

The price of that usefulness is complexity. Langfuse is priced like infrastructure, governed like infrastructure, and best adopted by teams that already think in those terms. For those buyers, it is an easy recommendation. For everyone else, it is a reminder that observability tools become valuable only after the product is serious enough to need them.