OpenSRE is a framework for AI site reliability engineering (SRE) agents that investigate production incidents on your own infrastructure. It connects to more than 60 tools across observability, cloud infrastructure, databases, data platforms, and incident management, and answers questions such as why a service is slow with an evidence-linked response. The project is in public alpha.
When you ask a question or an alert arrives, the agent fetches correlated logs, metrics, traces, and recent deploys, optionally masks sensitive identifiers before calling an external LLM, tests hypotheses in a tool-calling loop across connected systems, and answers with linked evidence. It suggests next steps, can execute remediation actions, and posts a summary to Slack, PagerDuty, or Telegram.
The project also builds an open reinforcement learning environment for incident response: end-to-end tests against realistic, cloud-backed production failures, intended as a benchmark and training ground for AI SRE agents in the way SWE-bench serves coding agents.
You use it through an interactive shell with slash commands for sessions, costs, and integrations, a headless CLI for scripts and CI jobs, or a Python API. Deployment options are a managed OpenSRE Cloud gateway, a Telegram gateway on an AWS server image, or a self-hosted container on Railway, ECS, or Vercel.
| Category | Examples |
|---|---|
| LLM providers | Anthropic, OpenAI, OpenAI Codex, Ollama, Google Gemini, OpenRouter, NVIDIA NIM, Bedrock |
| Observability | Grafana, Datadog, Honeycomb, CloudWatch, Sentry, Splunk, New Relic |
| Infrastructure | Kubernetes, AWS, Azure, ArgoCD, Helm |
| Databases | PostgreSQL, MySQL, MongoDB, ClickHouse, Snowflake, Redis, Supabase |
| Incident management | PagerDuty, Opsgenie, Jira, Alertmanager, incident.io, ServiceNow |
| Communication | Slack, Discord, Telegram, WhatsApp |
Features
- Evidence-driven diagnosis: correlates logs, metrics, traces, deploys, and config, and links every conclusion to the data behind it
- Runbook-aware reasoning: reads your runbooks and applies them during an investigation
- Identifier masking: reversibly redacts pods, clusters, and account IDs before external LLM calls and restores them in the output
- Remediation: suggests next steps and optionally executes remediation actions
- Incident reporting: posts summaries to Slack, PagerDuty, or Telegram
- Session cost and history: per-session token tracking and resumable shell sessions
- Local agent fleet: monitors Claude Code, Cursor, Codex, and other coding agents running on your machine
- Model choice: Anthropic, OpenAI, Codex, Ollama, Gemini, OpenRouter, TrustedRouter, NVIDIA NIM, and Bedrock
- Headless and embedded use: a non-interactive CLI for scripts and CI, and a Python API for in-process sessions
- Incident test environment: end-to-end tests over realistic cloud-backed failure scenarios for training and evaluating agents
