An AI SRE that traces topology and asks before it acts

What it does Ongrid is a self-hosted operations agent written in Go. When an alert fires, it spawns an investigator that collects metrics, logs, traces, and topology context, then produces a root-cause analysis with concrete next steps. Operators interact with it through Slack, Telegram, Lark, DingTalk, or WeCom, and risky actions are gated behind approval workflows.

The interesting bit The architecture avoids inbound ports entirely: edge nodes dial out to the central server, so hosts don’t need 22, 80, or 443 open. It also brings a browser-based reverse-tunnel SSH shell that works without keys or jump boxes, and routes queries to multiple LLM providers (Anthropic, OpenAI, DeepSeek, Gemini, GLM, Kimi) with hot routing.

Key highlights

• Coordinator dispatches tasks to specialist sub-agents (SRE, network, database) while a reviewer enforces policy boundaries.
• Built-in observability stack: Prometheus, Loki, Tempo, and Grafana are wired in, and the agent writes its own queries against them.
• RAG-backed knowledge vault indexes runbooks, incident history, and repositories so both humans and agents search the same context.
• Supports external MCP servers to extend the tool catalog available to agents and workflows.
• Every host inspection call runs through a read-only bash sandbox and is audited.

Verdict Teams drowning in alert noise and running Prometheus-family observability will get the most value; if your stack is entirely SaaS-hosted and you lack infrastructure to self-host, this is not your tool.