ITOps Agent Platform
An open-source multi-agent IT operations platform that turns alerts into root-cause diagnosis, approval, auto-remediation and verification — one closed loop.
Evidence shows claims of non-root containers, a 7-category dangerous-command filter, HITL approval closed loop, AES-256-GCM encryption and a 3-tier JWT secret strategy (documented concretely in SECURITY.md/README), but the supplied files contain no implementation of the command filter, approval engine, or SSH execution — least privilege, external-effects protection and rollback are 'asserted without code'. README promises only 'auto verification, failure alerting', with no rollback mechanism shown, while the product SSH-executes AI-generated commands and operates Docker/K8s/vSphere: large blast radius, thin evidence. Data-flow: multiple cloud LLM providers coexist with 'local AI, data stays in-domain', but nothing clarifies which alerts/logs/credentials are sent to LLMs. No dependency audit or SCA evidence; the recommended install path is a curl|sh remote script, a high-risk default. Deducted for: key security claims lacking corresponding source code.
CI is thorough (lint, tsc, architecture checks, frontend/backend tests, Docker build verification); auth integration tests use a real in-memory SQLite DB covering login/refresh/blacklist/audit persistence; error responses are structured with clear 400/401 semantics — failure_messages earns 2. However the README itself admits 'recent Docker images are transitional, unverified, with messy dependencies, not recommended' while simultaneously recommending a one-click deploy for a '5-minute complete loop' — a direct self-contradiction. Multi-model fallback and circuit breakers are declared only, with no test evidence. Deducted for: install path conflicting with stability statements.
Audience and scenarios are well covered: per-role pain-point tables (ops engineer/SRE/IT lead/SMB/security) and four clear usage paths — a 3. Environment fit is supported by Docker deploy and cloud/local model options. But capability boundaries are badly distorted: 'ultimate form of IT ops', 'marginal cost near zero', 'AI carries 80%+ of workload' are unfalsifiable marketing claims; no statement of AI misdiagnosis limits, unsuitable failure types, or minimum human-supervision requirements. Trigger precision: webhook signature modes are documented, but alert dedup/correlation accuracy has zero verification evidence. Deducted for: vision narrative burying real boundaries.
Information architecture is excellent: 7-language README, doc navigation table, clear project structure — a 3. Known limitations are honestly stated (transitional refactor images). Maintenance responsibility is clear (SECURITY.md with 48h ack / 30-day fix SLA, named author). But licensing has substantive problems: LICENSE is standard MPL-2.0, yet the README's 2026-05-27 notice adds restrictions MPL-2.0 does not contain ('no closed-source repackaging for sale', 'no competing SaaS'), and the MIT-to-MPL dual-era boundary is absent from LICENSE; metadata is NOASSERTION — license scores 1. No CHANGELOG file, only a badge and version 3.0.5. Examples/FAQ are thin: no concrete config examples. Install notes are docked for the curl|sh pattern plus the unstable-image warning.
Output usability is evidenced: structured JSON remediation commands, Markdown reports, natural-language diagnostics described concretely; tests confirm a uniform response format — a 2. Marginal value: the alert→diagnose→fix→approve→verify loop and the 'each open-source tool covers one segment' competitive logic are coherent and plausibly differentiated. Cost-benefit only 1: LLM API costs, self-hosting resource needs are entirely unmentioned, transitional instability raises trial cost, and '$0 annual cost' ignores model spend and ops effort. Deducted for: incomplete TCO narrative.
Claim traceability is weak: figures like 68 routes / 72 services / 63 pages / 32 migrations / 21 knowledge entries cannot be checked against the supplied files; '7-category command safety engine' and 'bank-grade AES-256' have no corresponding code shown. Cross-source corroboration is limited: README security claims match SECURITY.md, and auth integration tests corroborate login/blacklist behavior — but that covers only a small slice of the product. Fact/inference separation is poor: Gartner/CNCF trends, the $400B market, GitLab analogies, and 'the only open-source project to engineer the full-loop' are speculation interleaved with functional description without sources. Deducted for: numeric and marketing claims lacking a checkable evidence chain.
- The README adds restrictions ('no closed-source resale', 'no competing SaaS') that are not part of standard MPL-2.0 and conflict with the LICENSE file; obtain legal review before commercial use. Code is split across an MIT era (pre-2026-05-27) and MPL-2.0 era; verify boundaries commit by commit.
- The project itself describes current Docker images as 'transitional, unverified, messy dependencies' — do not use in production; never blindly run the curl|sh one-liner, audit deploy.sh first.
- Auto-remediation SSH-executes AI-generated commands and operates Docker/K8s/vSphere; no implementation code for the command filter or approval engine appears in the reviewed evidence — audit manually and rehearse in an isolated environment before production use.
- 'Data stays in-domain' only holds with local models; with cloud LLMs (Doubao/DeepSeek/OpenAI etc.), alert content, logs and potentially credentials leave your network — map the data flows explicitly.
- No evidence of a rollback mechanism on failed remediation (only failure alerting); prepare your own rollback plan before enabling auto-remediation on critical systems.
- Default account admin/admin: change the password immediately on first login and configure JWT_SECRET and WEBHOOK_VERIFY_ENABLED=true.
What does this agent do, and when should you use it?
ITOps Agent Platform (GitHub: qinshihu/itops-agent-platform) positions itself as China's first enterprise multi-agent IT operations platform, combining capabilities comparable to PagerDuty, Rundeck and Portainer, with Docker-based self-hosted deployment. The stack is Node.js + Express + TypeScript on the backend, React 18 + Vite + Ant Design on the frontend, and SQLite (better-sqlite3, WAL mode, AES-256 encrypted) for storage, spanning 12 preset agents, 68 API routes, 72 services and 63 frontend pages. The core execution loop receives Prometheus/Zabbix alert webhooks, performs AI root-cause analysis, generates remediation commands, routes them through human-in-the-loop approval pushed to WeCom/DingTalk, executes fixes over SSH, then verifies and reports automatically. Beyond alert remediation, the platform offers drag-and-drop workflow orchestration, unified Docker/VMware/Proxmox/Kubernetes management, IP subnet and VLAN management, a 3D data-center digital twin, knowledge-base RAG, and scheduled inspections. Its AI model pool supports Doubao, Qwen, DeepSeek, Zhipu, OpenAI, Claude and local models via Ollama, with primary/backup failover. Caveat: the project is undergoing a major 4A/DDD refactor; recent transitional code and images are explicitly flagged as unverified, and the README recommends that ordinary users wait for a stable release.
The platform ingests Prometheus and Zabbix alerts via webhook; an AlertProcessor state machine handles reception, filtering, correlation and noise reduction. Multi-agent AI (12 preset agents including alert handling, fault diagnosis, log analysis and command generation) produces natural-language root-cause reports and structured JSON remediation commands. Commands pass through an SSH command safety engine (7 categories of dangerous-command policies, role-based blocking), are pushed for approval to WeCom or DingTalk, and after one-tap mobile approval are executed automatically over ssh2, with automatic result verification and Markdown report generation. The automation domain's WorkflowEngine state machine supports drag-and-drop workflow authoring and Cron-scheduled inspections; the infrastructure domain manages multi-host Docker via Dockerode, VMware vSphere/Proxmox VE/KVM via dedicated adapters, and Kubernetes via kubeconfig cluster import. The network domain includes adapters for 17 vendors, SNMP topology discovery, and a data-center module providing rack modeling with Three.js WebGL 3D real-time monitoring. An MCP module bridges agent tools through a 6-layer securityGate, and a knowledge base injects context via RAG semantic retrieval. All data stays in local SQLite with encrypted backup and CSV/JSON import-export.
- On-call engineers: instead of waking up to SSH into a server at 3 AM, the AI diagnoses the root cause of a CPU-99% alert and the engineer approves a one-tap fix from a phone — about 3 minutes end to end
- SRE/DevOps teams: consolidate PagerDuty + Rundeck + Portainer-style tooling into one self-hosted platform with a closed alert-to-remediation loop
- SMB IT teams: cannot afford PagerDuty/ServiceNow pricing of $50,000+/year and need a free, open-source alternative that keeps data on-premises
- Security and compliance teams: require HITL approval on every remediation, full audit trails, and dangerous-command blocking for auditability
- Data-sensitive and xinchuang (domestic-stack) enterprises: connect local models via Ollama/LM Studio/vLLM so AI inference never leaves the domain
- Container and virtualization ops: need unified management of multi-host Docker, VMware vSphere, Proxmox VE and K8s with scheduled automated inspections
What are this agent's strengths and limitations?
- The only open-source AIOps project that claims an engineered full-loop of alert → diagnosis → decision → approval → execution → verification, with a 7-node state machine
- 12 cooperating agents rather than single API calls, combined with auditable reasoning chains, a command safety engine and HITL approval — production-oriented by design
- Unified AI model pool across providers (Doubao/DeepSeek/Qwen/OpenAI/Ollama etc.) with primary/backup failover and per-provider circuit breakers, including fully local models
- Broad coverage: container/VM/K8s management, IP/VLAN management, 3D data-center digital twin, workflow orchestration and knowledge-base RAG in a single platform
- Documented engineering maturity: 32 schema migrations, 68 API routes, 72 services, CI/CD workflows — beyond demo-level projects
- The project is mid-refactor (4A/DDD); recent code and Docker images are transitional, unverified, and self-described as having messy dependencies — users must wait for a stable version
- License changed: MIT for pre-2026-05-27 code, MPL-2.0 thereafter, with explicit bans on closed-source repackaging for resale and on building a competing SaaS — commercial use requires careful compliance review
- Runtime depends on Node.js 22 and a single-node SQLite database; scalability and multi-node deployment beyond that are not evidenced in the README
- High complexity (63 pages, 72 services, 60 migrations) implies significant learning and operational overhead when troubleshooting
- GitHub lists the license as NOASSERTION; the actual MIT/MPL-2.0 mix needs independent verification before adoption
How do you install or deploy this agent?
One-line script deployment (Linux/Mac, Docker required):
curl -sL https://gitee.com/IT_Oline/itops-agent-platform/raw/main/deploy.sh -o deploy.sh && chmod +x deploy.sh && ./deploy.shOn Windows PowerShell, run .\deploy.ps1.
Or via Docker Compose:
docker compose up -d --buildAfter deployment the frontend is at http://localhost:8080 and health checks at http://localhost:3001/health. For local development, enter the local-dev directory and run start-dev.sh (start-dev.bat on Windows); frontend on port 3000, backend on port 3001. Important: the project is in a transitional 4A/DDD refactor; the README explicitly warns that recent Docker images are unverified with messy dependencies and should not be pulled directly — ordinary users should wait for a stable release.
How do you use this agent?
- Open http://localhost:8080 after deployment and log in with the default admin/admin account (password change is forced on first login). 2. Add a server; the system auto-discovers containers and resources on the host. 3. Configure an AI provider (Doubao/Qwen/DeepSeek/OpenAI/Ollama etc.) in settings and wire up alert webhooks; fire a test alert and watch the AI analyze it. 4. Click "auto-fix" and approve from the WeCom/DingTalk mobile push to complete the full alert → diagnosis → approval → execution → verification loop. 5. Use the workflow editor to drag-and-drop agents and approval nodes, and configure Cron schedules for multi-server inspections with automatic reports.
How does this agent compare with similar options?
The README compares the project to Grafana OnCall, Portainer, Uptime Kuma, Rundeck and Coolify: those tools each cover one slice (alerting, containers, execution) and none offer multi-agent AI, auto-remediation closed loops or HITL approval. Against commercial suites (PagerDuty + Rundeck, ServiceNow ITOM), the commercial options cost $50,000–$100,000+/year and are cloud-forced SaaS, while this platform is free and 100% self-hosted — though it cannot match their maturity and enterprise support.