// the find
Tracer-Cloud/opensre
Build your own AI SRE agents. The open source toolkit for the AI era.
OpenSRE is a pre-alpha Python framework for building AI-driven incident response agents that connect to your existing observability stack (Grafana, Datadog, CloudWatch, etc.) and reason across logs, metrics, and traces to suggest or execute remediation. The stated goal is as much about building a benchmark and RL training environment for AI SRE as it is about shipping a production tool. Teams tired of context-switching between 6 dashboards during an outage are the target audience.
The synthetic incident simulation suite with scored RCA accuracy checks is the most interesting thing here — it's the missing piece that makes AI SRE trainable rather than just vibe-driven. The 60+ integration list is genuinely broad and covers the actual tools engineers run (Kubernetes, CloudWatch, PagerDuty, Kafka) rather than toy connectors. Supporting multiple LLM backends (Anthropic, OpenAI, Ollama, Gemini, Bedrock) means you're not locked in, and local Ollama support matters for teams with data residency constraints. The routing test harness using YAML scenario contracts is a solid pattern for testing agent decision-making deterministically.
It's pre-alpha and the benchmark table in the README is literally empty — 'No benchmark results yet' — which undercuts the core claim that this is a training environment. The dependency on a running FastAPI server plus Redis plus Postgres for hosted deployments adds real operational weight for what's pitched as a self-service SRE tool. Telemetry is opt-out (PostHog + Sentry), which is a trust problem for teams whose whole pitch for self-hosting is data control. The codebase directory structure is deep and the agent architecture has already been rewritten once (the README links to AGENT_ARCHITECTURE.md explaining the removal of 'old graph and chain framework layers'), which signals churn in a project that hasn't shipped a stable release.