ITOps Agent Platform is a self-hosted IT-operations platform that runs the path from alert to verified fix with a set of cooperating AI agents. An alert arriving by webhook from Prometheus or Zabbix is de-duplicated and correlated, analysed into a natural-language root-cause report, turned into a structured remediation command with a risk assessment, pushed to WeCom or DingTalk for human approval, executed over SSH once approved, verified, and written up as a report.
Twelve preset agents cover alert handling, fault diagnosis, log analysis, system inspection, change execution, document generation, compliance checking, command execution, scheduled inspection, command generation, network inspection, and database operations. A knowledge base with retrieval-augmented generation injects retrieved context into the model call, and an agent tool registry alongside an MCP (Model Context Protocol) layer with a multi-stage security gate exposes platform and external tools to those agents.
Beyond the alert loop the platform manages infrastructure directly: multi-host Docker, VMware vSphere, Proxmox VE and KVM virtual machines, Kubernetes clusters imported by kubeconfig, IP subnet and VLAN planning, data-centre rack and power modelling with a WebGL 3D room view, SNMP topology discovery across seventeen vendor adapters, and a browser SSH terminal. Workflows are composed by drag-and-drop with serial, parallel, and conditional branches and fired on a cron schedule.
Security controls include AES-256-GCM encryption of stored credentials, JWT access and refresh tokens with role-based access control, an SSH command filter covering seven classes of dangerous command applied per role, login lockout, and a full audit log. The model pool manages a primary-to-fallback chain with a separate circuit breaker per provider:
| Type | Providers | Access |
|---|---|---|
| Domestic cloud | Volcano Engine Doubao, Alibaba Qwen, DeepSeek, Zhipu GLM-4, Moonshot Kimi, Baidu Ernie, Yi, Baichuan | native API or OpenAI-compatible |
| International cloud | OpenAI GPT-4o, Anthropic Claude | native API |
| Local deployment | Ollama, LM Studio, vLLM | OpenAI-compatible, data stays on-premises |
The codebase is being restructured onto a 4A architecture with domain-driven design, replacing parts incrementally. Recent commits and the Docker images built from them are described as transitional and unverified, so a stable release is the recommended path for general use, while developers who can debug and modify the source can build from it directly.
Features
- Twelve preset agents: alert handling, diagnosis, log analysis, inspection, change execution, document generation, compliance, command execution and generation, network inspection, and database operations
- Alert-to-fix loop: webhook intake, de-duplication and correlation, root-cause report, generated commands, approval, SSH execution, verification, and report
- Human approval: approval nodes in workflows pushed to WeCom or DingTalk for sign-off from a phone
- Workflow editor: drag-and-drop serial, parallel, and conditional branches with ten preset templates and cron triggers
- Container and virtual machine management: multi-host Docker with Compose, plus VMware vSphere, Proxmox VE, and KVM with snapshots and live migration
- Kubernetes: clusters imported by kubeconfig, with Pod, Deployment, Service, and Node lifecycle
- Network and data centre: IP subnet and VLAN planning, rack, PDU and UPS modelling, a 3D room view, and SNMP topology discovery across 17 vendor adapters
- Knowledge base: semantic retrieval that injects context into the model call
- Command safety: seven classes of dangerous command filtered by role, with a full audit trail
- Model pool: primary-to-fallback chains with per-provider circuit breakers and connectivity tests
- Cost and scaling: container and virtual machine cost estimates, and metric-driven autoscaling with cooldown windows
