Both trace agents, run experiments, score production traffic and version prompts. Langfuse is the open-source option you can self-host for free, with evaluators that run on its own servers; Logfire traces the whole application, not only model calls, and adds a gateway that can block model requests. If self-hosting without a contract is a requirement, that settles it; otherwise compare how much of your system you want in one trace.[3][4]
Langfuse is an open-source LLM engineering platform focused on model and agent traces, evals and prompt versioning, while Logfire is Pydantic's OpenTelemetry observability platform that traces the whole application and adds an AI gateway. Langfuse is MIT-licensed apart from add-ons and free to self-host, and ClickHouse acquired it in January 2026; Logfire's platform is closed source and self-hosting is Enterprise only. The names are similar, but the products answer different questions.
| Decision factor | Langfuse | Logfire |
|---|---|---|
| Ownership | Acquired by ClickHouse in January 2026; licence and self-hosting unchanged, per Langfuse | Pydantic, independent |
| Licence | MIT, except add-ons in the ee folders | SDKs MIT; platform closed source |
| Self-hosting | Free, via Docker Compose or Kubernetes | Enterprise only, on Kubernetes via Helm |
| Trace scope | LLM calls plus retrieval, embedding and API steps | Whole application: HTTP, database, logs, metrics, LLM and agents |
| Production evals | Evaluator rules run by Langfuse, with backfills up to six months | Evaluators run in your application in the background |
| Gateway | None; traces are exported out of the request path | Yes, with guardrails and spend caps on model calls |
| Free tier | Hobby: 50k units a month, 30 days of data | Personal: 10M records a month, 30-day retention |
| First paid tier | Core: $29 a month, 100k units | Team: $49 a month, 10M records, 5 seats |
| Capability | Langfuse | Logfire | Prefactor |
|---|---|---|---|
| Observe | |||
| LLM call tracing | ✓ | ✓ | ✓ |
| Multi-step agent tracing (tool calls, nested spans) | ✓ | ✓ | ✓ |
| Cost tracking | ✓ | ✓ | ✓ |
| OpenTelemetry support | ✓ | ✓ | ✓ |
| Evaluate | |||
| Offline evals (datasets, experiments) | ✓ | ✓ | ✓ |
| Online evals on production traffic | ✓ | ✓ | ✓ |
| LLM-as-a-judge scoring | ✓ | ✓ | ✓ |
| Code-based evaluators | ✓ | ✓ | ✓ |
| Human review and annotation | ✓ | Partial | ✓ |
| Prompt versioning and playground | ✓ | ✓ | — |
| Act on findings | |||
| Runtime controls on live agent actions | — | Partial | ✓ |
| Gateway or proxy in the request path | — | ✓ | — |
| Licence and deployment | |||
| Open source | ✓ | Partial | Partial |
| Self-hosting | ✓ | Partial | Partial |
Langfuse publishes its platform under the MIT licence, apart from add-on features in licence-keyed ee folders. Self-hosting is free on Docker Compose for testing or Kubernetes, AWS, Azure or GCP for production, on Postgres, ClickHouse, Redis or Valkey and S3-compatible storage. ClickHouse acquired Langfuse in January 2026, and Langfuse said its roadmap and licence would not change. Releases have kept coming since, up to v4.64.0 on 19 October 2026. One date to note for cloud users: Langfuse Cloud becomes v4-only on 16 November 2026, which may require migration work.
Logfire's platform is closed source. Its SDKs and pydantic-evals are MIT-licensed, but running the platform yourself needs the Enterprise plan, a Kubernetes cluster via Helm, PostgreSQL 16 or later and object storage; a dedicated single-tenant deployment in any GCP region is also offered. If traces must stay on your own servers without a commercial agreement, Langfuse is the one of these two that allows it.
Langfuse traces LLM calls and the non-LLM steps around them, such as retrieval, embeddings and API calls, with nested observations, sessions for multi-turn work and an agent graph view added in July 2026. It acts as an OTLP backend over HTTP (gRPC is not supported) and names OpenLLMetry and OpenLIT as compatible. Its scope is the LLM application.
Logfire's scope is the whole service. Its SDKs for Python, JavaScript and TypeScript, and Rust are OpenTelemetry wrappers that record incoming HTTP requests, database queries, logs and metrics in the same trace as model and tool calls, with Agents and LLMs views on top, and any other OpenTelemetry SDK can send data. Teams without a separate APM tool get both jobs from one product; teams that already have one may prefer Langfuse's narrower focus.
On prompts the two are near parity. Langfuse versions prompts, deploys a version by label without a code change, links each version to the traces that used it and compares metrics across versions, with a playground. Logfire stores immutable versions, tests them against saved scenarios and datasets, promotes them with labels and rollout targeting, and fetches them at runtime through its SDK. Both also track token cost: Langfuse per generation, ingested or inferred from model definitions with custom prices, and Logfire as a queryable trace column where token data exists.
Langfuse runs evaluators on its side. LLM-as-a-judge evaluators and code evaluators, written in Python or TypeScript in the UI since May 2026, score live observations at a sampling rate, and since September 2026 backfills can score up to six months of past data. A GitHub Action runs experiments in CI and fails below a threshold. Annotation queues let reviewers score traces, observations or sessions.
Logfire's live evaluations run in your application through pydantic-evals, in the background at a sample rate, and report as OpenTelemetry events. Its run annotations are in beta and annotation queues are early access.
At run time, Langfuse blocks nothing; its docs say protection comes from separate libraries and Langfuse evaluates them afterwards. Logfire's gateway can flag, redact or block model calls and cap spend, for traffic routed through it. Neither covers an agent's tool actions.
As of 20 October 2026, Langfuse charges per unit, where a unit is a trace, observation or score: Hobby is free for 50k units a month with 30 days of data, Core is $29 a month for 100k units then $8 per 100k, falling to $6 per 100k above 50M, and Pro is $199 a month. Logfire charges per record, where a record is a span, log or metric: Personal is free for 10M records, Team is $49 a month for five seats, Growth is $249 a month, and extra records cost $2 per million.
The units do not convert directly, and Logfire counts non-LLM spans too, so model both on a week of your own traffic using Langfuse's pricing and Logfire's pricing. Self-hosted Langfuse has no subscription but carries your infrastructure cost.
We sell the layer this section describes. Read it with that in mind.
Stated plainly, since Prefactor publishes this page: Prefactor is an agent evaluation platform that scores each step of a run and acts on what it finds through runtime policies on tool calls and data access, approval holds for risky actions and a kill switch for a run, an agent or a team. Langfuse evaluates after the fact, and Logfire's gateway checks model calls only; neither covers tool actions or approvals. Prefactor has no prompt playground or prompt registry and its platform is hosted, so if free self-hosting is what you need, Langfuse is the better pick. See the full comparison of 15 platforms.
Reviewed against public product and documentation pages on October 20, 2026. If a vendor has changed a feature, product name, or positioning since then, send a correction and we will update it. Numbered source links in the page body point to the ordered sources below.
Prefactor scores each step of every agent run and can hold or stop an action when a score says it should not happen. Start on the free tier: 5,000 spans a month, no credit card required.
Prefactor helps teams observe, evaluate, and improve their AI agents in production, across every framework and provider.