Ongrid
An ops AI Agent that understands your infrastructure, finds the root cause, and fixes it — right from Slack, Telegram, Lark or DingTalk.
README asserts read-only host tools, audit trails, approval write gate, and zero inbound ports, but no core implementation code is included to verify these; the safety claims remain marketing-level, so least_privilege, user_confirmation and external_effects get 1. Data flow (edge dial-out, MCP trust boundaries, registry allow-list) is described concretely: 2. Sensitive data handling only mentions DPAPI and credential handling without detail: 1. Dependencies are fully pinned with solid CI gates, but no vulnerability scanning (e.g., govulncheck) is evidenced: 2. Upgrade scripts exist but rollback is undocumented: 1. AGPL-3.0 plus trademark carve-out is clear: 2.
CI covers Go build/vet/race tests, Windows cross-compile, web tests, and deployment script checks, with motivated comments and consistent scope: 2. Pinned dependencies and a China CDN mirror support availability: 2. No evidence of failure-message quality: 1.
Clearly targeted at ops/SRE teams across alerts, RCA, Kubernetes, and network devices, with a documentation map: 2. Skills catalog, MCP servers, and tool-class inventories make capability boundaries visible: 2. No evidence on alert-trigger precision or false-trigger controls: 1. Supported environments (Ubuntu 22.04+/Debian 12+/RHEL 9, amd64/arm64) are explicit: 2.
README structure is complete, but substantial content lives on the external ongrid.cloud docs site, limiting in-repo information: 2. Install steps cover both architectures plus a CDN mirror: 2. Naming (ongrid/edge/pack) is consistent: 2. Product tour screenshots and a demo GIF exist, but no FAQ or known-limitations list: examples 2, limitations 1. Full LICENSE and trademark boundary: 3. Version badge and release artifacts (v0.15.2) exist but no in-repo CHANGELOG: 2. SECURITY.md defines private reporting and response times: 2.
RCA reports and approval UI screenshots suggest outputs are designed for human review, but this is static image evidence: 2. The value proposition of combining observability, topology, and agent execution is clear: 2. No quantified cost/benefit (LLM spend, hosting, edge footprint): 1.
Core claims (read-only tool counts, auditing, write gate) cannot be traced to implementation in the provided files: 1. Limited cross-source corroboration within the repo; no independent corroboration: 1. Marketing and factual content are mostly distinguishable but not systematically separated (badges mixed with feature claims): 1.
- The product performs remote execution and reverse-tunnel shell access, yet no implementation code in the evidence verifies the claimed "read-only tools" or "approval write gate"; audit the core code before deploying.
- Security claims (zero inbound ports, full auditing, read-only sandbox) are README assertions only; do not grant write permissions in production based on them.
- No dependency vulnerability scanning (govulncheck/dependabot-style) and no in-repo CHANGELOG are evidenced; verify upgrade and rollback paths yourself.
- Publisher is unverified and much documentation lives off-repo; distinguish auditable repository evidence from marketing content.
What does this agent do, and when should you use it?
Ongrid (ongridio/ongrid) is an open-source, self-hosted ops AI Agent written in Go with a TypeScript/React frontend. It ingests metrics, logs, traces and topology, and on alert triggers spawns an RCA worker that correlates evidence and pins the "why" to a source-code line. A coordinator Agent dispatches to SRE, network and database specialist sub-agents; an Edge component dials out to hosts, enabling browser SSH, read-only inspection tools and an approval write gate. The platform ships a built-in Prometheus + Loki + Tempo + Grafana observability stack, hot-swappable model routing (Anthropic, OpenAI, GLM, DeepSeek, Gemini, Kimi), plus a Workflow Builder, MCP server registry, knowledge vault and artifacts center. Hosts require zero inbound ports, and every tool call is audited.
The execution flow: an alert or operator question triggers the incident investigator, which spawns an RCA worker that walks the topology graph, computes blast radius, correlates metrics/logs/traces, and produces an evidence-backed root-cause report with concrete next steps. Agents reach hosts through the Edge component's outbound connection, run 26+ read-only inspection tools in a bash sandbox, and offer audited browser SSH without keys or a jumpbox; risky actions must pass the approval write gate. Operators work in a web console with monitoring dashboards, a topology map, a Workflow Builder (triggers, agents, tools, conditions, notifications), external MCP server registration, a Knowledge Vault (indexing runbooks, incident history and repos for RAG search shared by humans and agents), and an Artifacts Center. It also enrolls Kubernetes clusters for workload/event inspection and upgrade management, and discovers network-neighbor devices from Edge hosts with SNMP verification and interface polling.
- An SRE team receives a Prometheus alert at night; the agent auto-investigates and writes the root cause back to the Telegram group without manual dashboard-hunting.
- An on-call engineer needs shell access to a host and connects via the audited, key-free browser SSH reverse tunnel.
- A platform team enrolls multiple Kubernetes clusters into Ongrid to inspect workloads and events and manage upgrades from one place.
- An ops lead encodes alert triage, investigation and notification into repeatable automations with the Workflow Builder.
- A team indexes runbooks and past incidents into the Knowledge Vault so humans and agents search the same operational context.
- Network operators discover switch neighbors from Edge hosts, verify via SNMP, and map host-to-network-device links.
What are this agent's strengths and limitations?
- Zero-inbound-port architecture: Edge dials out and hosts keep no open 22/80/443, sharply reducing attack surface versus jumpbox-based access.
- Alert-driven auto-investigation: alerts spawn an RCA worker that writes the root cause back to chat instead of just forwarding notifications.
- Evidence-backed RCA: walks topology, correlates metrics/logs/traces, and pins conclusions to a source-code line.
- No model lock-in: Anthropic, OpenAI, GLM, DeepSeek, Gemini and Kimi are supported with hot routing.
- Write gate for risky actions: production changes require approval, and all tool calls are audited.
- AGPLv3 licensing imposes copyleft constraints on redistribution and commercial integration; brand assets are excluded from AGPLv3 (see TRADEMARK.md).
- The one-command install targets Linux only (Ubuntu 22.04+, Debian 12+, RHEL/Rocky 9); no other OS or containerized deployment path is documented.
- Core value depends on wiring in Prometheus, Loki, Tempo, Grafana and Qdrant; teams with heterogeneous monitoring stacks face integration costs.
- Initial setup requires both a server and Edge components plus model API keys and IM channel credentials, adding operational overhead.
- Depth and maturity of specific capabilities (Kubernetes upgrade management, network device management) lack quantitative evidence (stars, production cases) in the README and need independent evaluation.
How do you install or deploy this agent?
On a server running Ubuntu 22.04+, Debian 12+, or RHEL/Rocky 9, download the release for your architecture and run the installer:
AMD64:
wget https://github.com/ongridio/ongrid/releases/download/v0.15.2/ongrid-v0.15.2-linux-amd64.tar.xz
tar -xf ongrid-v0.15.2-linux-amd64.tar.xz && cd ongrid-v0.15.2-linux-amd64
sudo ./install.shARM64:
wget https://github.com/ongridio/ongrid/releases/download/v0.15.2/ongrid-v0.15.2-linux-arm64.tar.xz
tar -xf ongrid-v0.15.2-linux-arm64.tar.xz && cd ongrid-v0.15.2-linux-arm64
sudo ./install.shIn mainland China, use the CDN mirror at https://ongrid.cloud/dl/ongrid-v0.15.2-linux-amd64.tar.xz if GitHub is slow. install.sh brings up the full stack in one command.
How do you use this agent?
After install, complete first-boot setup, configure a model provider (Anthropic / OpenAI / GLM / DeepSeek / Gemini / Kimi, with hot routing), and connect a two-way IM channel (Slack, Telegram, Larksuite, DingTalk or WeCom) with per-channel locale. Deploy Edge on target hosts — Edge dials out, so hosts need no inbound 22/80/443 ports — then ask questions in chat or let alerts trigger auto-investigation. Advanced usage: build automations in the Workflow Builder, register external MCP servers, import runbooks and repos into the Knowledge Vault, and enroll Kubernetes clusters. Full documentation is at ongrid.cloud/docs/get-started/introduction.