Ongrid is an open-source, self-hosted ops AI agent that reads metrics, logs, traces and topology to find the root cause of an incident and act on it from Slack, Telegram, Lark or DingTalk.

What it is

Ongrid is a Go-based AI agent for infrastructure operations, published under the AGPL-3.0 licence by ongridio and distributed as source on GitHub. It sits in the AIOps and ChatOps corner of the observability ecosystem: it wires Prometheus, Loki, Tempo and Grafana into one stack, connects to Kubernetes clusters and network devices, and places an LLM-driven reasoning layer on top, so an operator can ask a question in a chat channel and receive an evidence-backed answer. The project ships as a coordinator agent that dispatches work to specialist SRE, network and database sub-agents, backed by a RAG store over operational knowledge and source code.

The concrete problem it solves is the gap between an alert firing and someone understanding why. In a conventional setup, an alert fires, a human opens a dashboard, jumps onto a host through an SSH jumpbox with a key on disk, correlates metrics, logs and traces by hand, and then guesses at a cause. Ongrid replaces that manual jumpbox-and-key path with a reverse-tunnelled browser shell, and replaces ad-hoc chat triage with an investigator that spawns a root-cause worker when an alert fires and writes the cause back into the channel. Edge components dial out, so no inbound ports 22, 80 or 443 need to be open on managed hosts.

Key capabilities

• Coordinator plus specialist agents: the coordinator dispatches work to SRE, network and database sub-agents.
• Alert-driven auto-investigation: an investigator spawns an RCA worker and writes the cause back to chat.
• Root-cause analysis that walks topology, correlates metrics, logs and traces, and pins the cause to a source-code line, supported by RAG knowledge and code search.
• Zero inbound ports on managed hosts: edge dials out, and browser SSH runs over a reverse tunnel with no keys, no jumpbox and full auditing.
• Read-only host tools: a bash sandbox plus 26 or more inspection tools, with every call audited.
• Built-in observability: Prometheus, Loki, Tempo and Grafana are wired together, and the agent writes the queries.
• Bring your own model: Anthropic, OpenAI, GLM, DeepSeek, Gemini and Kimi with hot routing, plus two-way Slack, Telegram, Larksuite, DingTalk and WeCom channels with per-channel locale.

Who uses it and how

• On-call SRE teams that already live in Slack, Telegram, Larksuite, DingTalk or WeCom and want incident investigation to happen in the channel rather than in a separate dashboard.
• Kubernetes platform teams that enroll clusters through Edge, inspect workloads and events, manage upgrades, and mirror Kubernetes resources into topology.
• Network operations teams that discover neighbors from Edge hosts, verify them with SNMP, poll interfaces, and map host-to-network-device links.
• Environments where hosts must not expose inbound ports, since the edge component dials out and browser SSH uses a reverse tunnel with auditing.
• Self-hosters who want the full observability stack and their own model credentials in one deployment; operators in mainland China can pull from the ongrid.cloud CDN mirror when GitHub is slow.

Getting started

Download the release tarball for the server architecture, ongrid-v0.16.0-linux-amd64.tar.xz or ongrid-v0.16.0-linux-arm64.tar.xz, extract it, and run sudo ./install.sh, which brings up the full stack in one command. It targets Ubuntu 22.04+, Debian 12+ and RHEL/Rocky 9.

How it compares

These facts do not name a set of paid products that Ongrid replaces, and they name no directly comparable tools either. It therefore stands alone in this registry, and any comparison would have to be drawn from sources outside the material provided.

When to use it — and when not to

A self-hoster must operate the bundled Prometheus, Loki, Tempo and Grafana stack, run edge agents on managed hosts, and supply model credentials for one of the supported providers, which brings its own cost and data-flow questions. The AGPL-3.0 licence is a real constraint for anyone embedding the code in a closed product, and the project carries 74 open issues, so teams expecting a polished, fully managed service should look elsewhere. Installation is also Linux-only through a tarball and install.sh; the facts mention no container image or compose file.

wget https://github.com/ongridio/ongrid/releases/download/v0.16.0/ongrid-v0.16.0-linux-amd64.tar.xz
tar -xf ongrid-v0.16.0-linux-amd64.tar.xz && cd ongrid-v0.16.0-linux-amd64
sudo ./install.sh

wget https://github.com/ongridio/ongrid/releases/download/v0.16.0/ongrid-v0.16.0-linux-arm64.tar.xz
tar -xf ongrid-v0.16.0-linux-arm64.tar.xz && cd ongrid-v0.16.0-linux-arm64
sudo ./install.sh

# AMD64
wget https://ongrid.cloud/dl/ongrid-v0.16.0-linux-amd64.tar.xz

# ARM64
wget https://ongrid.cloud/dl/ongrid-v0.16.0-linux-arm64.tar.xz