diff --git a/overview.mdx b/overview.mdx index dc8c2dd..ad087b8 100644 --- a/overview.mdx +++ b/overview.mdx @@ -3,36 +3,35 @@ title: Laminar sidebarTitle: Laminar --- -Laminar is an [**open-source**](https://github.com/lmnr-ai/lmnr) observability platform purpose-built for AI agents. You can trace every call to an LLM, tool execution and custom functions. +Laminar is an [**open-source**](https://github.com/lmnr-ai/lmnr), OpenTelemetry-native observability platform built for AI agents. **[Trace](/tracing/introduction) every LLM call, tool call, and subagent handoff, read runs as [transcripts](/platform/viewing-traces) instead of span trees, and [query, cluster, and alert](/signals/introduction) on the patterns that matter across thousands of traces.** -Laminar goes beyond tracing and is deeply focused on helping you properly use captured data and extract insights from it: -- **During development**, Laminar allows you to rollout and replay agents from different parts of the captured trace significantly speeding up the development process. - -- **In production**, with Laminar you can describe patterns and errors in natural language and track it across all your traces, to measure real behavior and user interactions with your agents. + + Multi-agent trace rendered as a transcript + - - Trace LLM calls, tool use, and custom functions. First-class support for **AI SDK, LangChain, Browser Use, Stagehand**, and more. + + First-class support for **Claude Agent SDK, OpenAI Agents SDK, Mastra, Pydantic AI, LangGraph, AI SDK, Browser Use, and Stagehand**. Every major LLM provider is auto-instrumented via OpenTelemetry. - Describe outcomes and failures in plain language. Laminar reads every trace and produces structured events you can query, cluster, and alert on. + Describe an outcome or failure in plain language (*did it get stuck, did the user give up, did checkout succeed*). Laminar reads every trace and produces structured events you can query, cluster, and alert on. - - Run evals against datasets locally or in CI. Catch regressions before they ship. + + Read an agent run as a conversation: auto-extracted `Input` block, inline previews on every LLM turn and tool call, subagents as collapsible cards you expand in place. - - Build evaluation datasets from production traces. + + Query every span, trace, and signal event directly with ClickHouse SQL from the in-app editor, the HTTP API, or a [Laminar MCP server](/platform/mcp) your agents can call. - - Replay any traced span. Swap prompts or models and compare results side by side. + + Run evals against datasets locally or in CI. Compare runs side by side and catch regressions before they ship. - - Realtime traces, full-text search, raw SQL access to all your data, and dashboards for exploring traces however you need to. + + Replay any traced span. Swap prompts or models and compare outputs side by side.