OpenSRE

OpenSRE is an open-source framework for building AI site-reliability agents that answer production questions and work through incidents on infrastructure you control. It targets the case where the evidence for a failure is scattered across logs, metrics, traces, runbooks, and chat threads, and it connects to the tools already producing that evidence rather than asking you to move the data.

A turn follows a fixed shape: fetch the relevant context and correlated logs, metrics, traces, and recent deploys; optionally mask sensitive identifiers such as pods, clusters, and account IDs before any external model call, restoring them in the output; reason across the connected systems in a tool-calling loop to test hypotheses; answer with evidence links; suggest next steps and optionally run remediation; and post a summary to Slack, PagerDuty, or Telegram.

The project also positions itself as training and evaluation ground for this class of agent. Its stated motivation is that coding agents got scalable training data and clear feedback from SWE-bench while incident response has no equivalent, so OpenSRE is building an open reinforcement-learning environment for agentic infrastructure incident response, with end-to-end tests over realistic, cloud-backed production failures.

You can drive it from an interactive shell with slash commands for session control, integrations, local agent-fleet monitoring, and reasoning effort; from a headless CLI for scripts and CI jobs; or in-process from Python through a session API. Deployment paths include an AWS EC2 gateway image with systemd and hosted options such as Railway, ECS, or Vercel using the repository Dockerfile. The project is in public alpha, so APIs and integrations may still change, and product analytics and error telemetry are opt-out.

Features

• Evidence-driven diagnosis: correlated analysis across logs, metrics, traces, deploys, and configuration, with each conclusion linked to its data
• Runbook-aware reasoning: the agent reads your runbooks and applies them during an investigation
• Reversible identifier masking: pods, clusters, and account IDs are redacted before external model calls and restored in the output
• Three interfaces: an interactive REPL, a headless CLI for scripts and CI, and a Python session API
• Local agent fleet: monitors Claude Code, Cursor, Codex, and other coding agents running on your machine
• Model flexibility: Anthropic, OpenAI, Codex, Ollama, Gemini, OpenRouter, TrustedRouter, NVIDIA NIM, and Bedrock
• 60+ integrations: observability, cloud infrastructure, databases, data platforms, dev tools, incident management, and communication
• Protocol support: MCP and ACP alongside the direct integrations
• Session cost and history: per-session token tracking and resumable sessions
• End-to-end test suite: cloud-backed failure scenarios kept separate from unit tests by catalog naming