Logfire

Key features

• OpenTelemetry-native full-stack tracing OpenTelemetry-native full-stack tracing: capture LLM calls in context with the HTTP requests, database queries, and business logic around them, across agents, services, and hosts on one correlated timeline. Any OTel-speaking language or framework can send data, so polyglot stacks land in the same view.[1][2]

• LLM panels with tokens and cost LLM panels with tokens and cost: inspect conversations and tool calls with token tracking, cost monitoring, a per-model LLMs view, and streaming support. The panels answer which model spent what on which task without exporting spans elsewhere.[1][2]

• PostgreSQL-compatible SQL queries PostgreSQL-compatible SQL queries: slice traces, logs, and metrics with familiar SQL instead of a proprietary query language. SQL-fluent developers filter, aggregate, and join telemetry the way they already query application data.[1][2]

• Streaming Live view Streaming Live view: watch traces and logs arrive in real time and drill into any span for its full detail. New instrumentation is verifiable within minutes of the first trace rather than after index delays.[1]

• Dashboards, metrics, and Slack alerts Dashboards, metrics, and Slack alerts: chart what matters, browse any metric by dimension in the metrics explorer without SQL, and get notified in Slack when it changes. Routine service, host, Kubernetes, and database telemetry ships with pre-built dashboards.[1]

• Code-first evals with live monitoring Code-first evals with live monitoring: define evaluations in Python with pydantic-evals, version-controlled beside tests, and manage hosted datasets, side-by-side experiments, and live grading of production traffic in the web UI. Evals cover tools, pipelines, and whole agent workflows, not only model calls.[1]

• Prompt management with rollouts Prompt management with rollouts: version prompts, run labeled rollouts, and iterate in a playground before shipping. Prompt changes become reviewable releases instead of scattered string edits.[1]

• AI Gateway with 40-plus integrations AI Gateway with 40-plus integrations: route models through unified key management with spending limits and provider failover, bring up to 3 provider credentials on Personal, and instrument FastAPI, Django, SQLAlchemy, HTTPX, and more with one line of setup each. A guided CLI wizard connects a directory to a project and verifies data arrival.[1][2][3]

Who will love it

• AI teams debugging agents in production will love it: when an agent fails, the trace shows the database timeout or rate limit behind the model error instead of only the LLM layer. Token and cost panels attribute spend per model and task, so prompt and model choices stay visibly tied to budget.[1][2]

• Python shops already on Pydantic AI, FastAPI, or OpenTelemetry will love it: three lines of code instrument the basics, the CLI wizard verifies arrival, and SQL-fluent developers query everything without learning another language.[1][2]

• Regulated or enterprise teams will love it: SOC2 Type II certification with HIPAA and self-hosted Enterprise deployment on their own Kubernetes keeps sensitive traces inside their boundary. The same features run in Cloud and self-hosted, so the deployment choice does not force a capability tradeoff.[1][2]

Who should avoid it

• Solo hackers avoiding any usage meter should avoid it: the free Personal tier hard-caps at 10 million records a month across 3 projects with 30-day retention, and ingestion pauses past the limit.[1][2]

• Non-technical teams should avoid it: instrumentation, SQL querying, and eval definitions all assume developers own the integration. The visual UI inspects results but cannot replace code ownership for setup.[1][2]

• Teams needing free self-hosting should avoid it: self-hosting is an Enterprise sales arrangement, and the MIT license covers the SDKs rather than a runnable free platform. Open-source self-hosted alternatives in the catalogue cover that shape instead.[1]

Alternatives and compatibility