Latitude is an open-source observability tool for AI agents, with the core goal of enabling AI agents to self-heal. It organizes its functionality around four stages: "observe—understand—fix—verify."
Purpose: To help developers discover where AI agents fail in real-world operation, and to use coding agents to automatically complete fixes and regression verification.
Main capabilities:
- Observe: A single line of telemetry code captures all trace data, including multi-turn sessions, tool calls, and the complete execution path. It supports traces, sessions, tool-call observability, and OpenTelemetry ingest.
- Understand: Failed traces are automatically classified into trackable signals, with status, scale, and trends, making it easy to determine the type of issue and how often it occurs. It also provides flaggers, behaviours, semantic search, and annotation capabilities.
- Fix: Latitude dispatches the full context, sample traces, and deep links to coding agents (such as Claude Code and Cursor), which write the minimal correct fix and submit a PR. It supports Agent Dispatch, MCP server, Linear, and webhook dispatch.
- Verify: Fixes are replayed against real failed traces to ensure that fixed failures no longer recur. Related capabilities include regression datasets, monitors, and alerts.
How to use:
- You can paste a prompt directly into coding agents such as Claude Code, Cursor, Windsurf, Codex, and OpenCode to complete integration. You can also manually configure a TypeScript example (the documentation uses OpenAI as an example, which can be replaced with the LLM SDK already used by the application), and use capture() at the request, conversation, or agent boundary to add user ID, session ID, tags, or metadata. Python and any runtime compatible with OpenTelemetry are also supported.
- Latitude is model-vendor agnostic, and its telemetry works out of the box with most model providers and frameworks (OpenAI, Anthropic, Bedrock, Vercel AI SDK, LangChain, etc.) as well as any OTEL-compatible application.
- When building inside Claude Code, there is a dedicated package that can capture full session records as traces, suitable for terminals, desktop applications, and IDE extensions.
- All operations in the UI can also be performed from the coding agent side through the MCP server and CLI.
Deployment and access: It can be self-hosted on fully open infrastructure, and ready-to-use container images are available on Docker Hub. It supports single-machine Docker Compose (simple), Kubernetes + Helm (clustered, scalable and highly available), and one-click deployment based on Railway. The cloud plan can be used for free, includes 20K credits per month, 30-day data retention, and unlimited seats. You need to register at latitude.so to obtain an API key and project slug. The project uses the MIT license, welcomes contributions, and provides a Slack community.
Limitations: Specific model support scope, performance metrics, enterprise-grade features, and pricing details are not provided in the materials.
