Both watch agents in production, but they look at different things. Logfire puts agent traces in the same OpenTelemetry view as the rest of your application, with code-defined evals and a gateway; Raindrop tracks behaviours you describe across live traffic and groups recurring failures for you.[1][4]
Pydantic Logfire traces your whole application on OpenTelemetry, from HTTP and database calls to model and agent calls, and adds evals, prompt versioning and an AI gateway; Raindrop focuses on agent behaviour in production, classifying live events with custom signals and grouping failures into issues. Logfire bills by records, Raindrop by events. Both are independent companies.
| Decision factor | Logfire | Raindrop |
|---|---|---|
| Ownership | Pydantic, independent; Series A led by Sequoia in 2024 | Independent; Series A led by CRV in September 2026 |
| What it traces | Whole application: HTTP, database, model and agent spans, plus logs and metrics | Agent events: messages, tool calls and decisions |
| Evals | pydantic-evals datasets and experiments; live evaluations run in your app | Signals classify every event; no hosted offline runner |
| Prompt tooling | Prompt versioning with labels and a Prompt Playground | None documented |
| Run-time controls | Partial: gateway guardrails and spend caps on routed model calls | None documented |
| Self-hosting | Enterprise only, Kubernetes via Helm | VPC deployment in beta for Enterprise and select partners |
| Free tier | Personal: 10M records a month, 1 seat, 30-day retention | Hobby: 1,000 events a month, 14-day retention |
| First paid tier | Team: $49 a month, 10M records, 5 seats | Pro: $299 a month plus $0.003 per event up to 1M events |
| Capability | Logfire | Raindrop | Prefactor |
|---|---|---|---|
| Observe | |||
| LLM call tracing | ✓ | ✓ | ✓ |
| Multi-step agent tracing (tool calls, nested spans) | ✓ | ✓ | ✓ |
| Cost tracking | ✓ | ✓ | ✓ |
| OpenTelemetry support | ✓ | ✓ | ✓ |
| Evaluate | |||
| Offline evals (datasets, experiments) | ✓ | Partial | ✓ |
| Online evals on production traffic | ✓ | ✓ | ✓ |
| LLM-as-a-judge scoring | ✓ | Partial | ✓ |
| Code-based evaluators | ✓ | Partial | ✓ |
| Human review and annotation | Partial | 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 | Partial |
| Self-hosting | Partial | Partial | Partial |
Logfire is an observability platform first. Its SDKs wrap OpenTelemetry, and any standard OpenTelemetry SDK, in Go, Java, .NET or Ruby, can send data to it. An agent's model calls, tool calls and retrieval steps appear next to the application's HTTP requests and database queries, with dedicated Agents and LLMs views. If a slow agent turn is really a slow database call, you see that in one trace.
Raindrop looks only at the agent. Traces follow each message, tool call and decision in a run, and the product is organised around what those runs mean: Signals mark behaviours you care about, Issue Detection groups recurring failures, and a Triage Agent investigates issues in the web app, in Slack or over MCP. It is not built to show your database latency.
So the first question is whether you want one tool for the whole service, or a dedicated view of agent behaviour alongside whatever you already use for the rest.
Logfire's evals come from pydantic-evals, an MIT-licensed Python library. You define datasets in code or host them in Logfire, run experiments against a baseline on demand or in CI, and score with LLM judges that record their reasoning or with plain code checks. Live Evaluations score production calls after they run: evaluators execute in the background inside your application at a sample rate you set, and Logfire aggregates the results. Run annotations for human review are in beta, and annotation queues are an early-access feature.
Raindrop asks you to describe a behaviour in plain language, label some sample events, and it builds a classifier that runs on every new event and can backfill 90 days of sampled history. You cannot supply your own judge prompt or model. Offline testing happens in Workshop, an open-source local debugger where a coding agent turns failures into evals, while hosted pull-request testing (Simulations) is in early access.
Engineers who want evals versioned next to their code will find Logfire's approach familiar. Teams who want behaviour tracking without writing judges will find Raindrop quicker to start.
The units differ, so a direct comparison needs your own numbers. As of 20 October 2026, Logfire counts records, meaning spans, logs and metrics. Personal is free for 10M records a month with one seat, three projects and 30-day retention; ingestion pauses at the cap. Team is $49 a month with 10M records and five seats, Growth is $249 a month with unlimited seats and up to 90-day retention, and extra records cost $2 per million.
Raindrop counts events. Hobby is free for 1,000 events a month with one custom signal and 14-day retention. Pro is $299 a month plus $0.003 per event up to 1M events, then $0.002, with unlimited retention. Logfire's records include the non-agent spans, logs and metrics it collects, so count those too when you estimate. See Logfire's pricing and Raindrop's plans.
Logfire includes the Pydantic AI Gateway, folded into Logfire in March 2026. Model calls routed through it can be observed, flagged, redacted or blocked by guardrail rules, and spend caps stop requests once a budget is spent. That covers model traffic only; tool actions outside the gateway are not checked. Raindrop has no gateway and documents no blocking: it analyses events after they arrive and alerts.
Cost tracking works differently as well. Logfire treats cost as a queryable column in traces wherever telemetry includes token data, and its gateway meters spend per project, user and key. Raindrop's Costs page breaks model spend down by model, provider and project, using captured usage and a versioned catalogue of public model rates.
Both platforms are closed source with open SDKs, and both reserve self-hosting for their largest customers. Logfire's self-hosted option is Enterprise only and needs Kubernetes, PostgreSQL 16 or later and object storage. Raindrop's VPC deployment is still in beta.
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 and a kill switch scoped to a run, an agent or a team. Logfire can block model calls that pass through its gateway but not tool actions, and Raindrop documents no run-time controls. Prefactor does not trace your wider application the way Logfire does, has no prompt playground or prompt registry, and its platform is hosted. 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.