XCP-ng AIops

Disclaimer: Community-maintained open-source project. Not affiliated with, endorsed by, or sponsored by Vates, the XCP-ng project, or the Xen Orchestra project. "XCP-ng", "Xen Orchestra", and "Xen" are trademarks of their owners. MIT licensed.

AI-powered XCP-ng operations via Xen Orchestra's REST API with a built-in governance harness — unified audit log, policy engine, token/runaway budget guard, undo-token recording, and descriptive risk tiers. Built for homelabs and small/self-hosted XCP-ng fleets that want an AI agent to triage VM health, storage pressure, backup failures, and patch posture — with every write audited, previewable, and (where honest) reversible. Self-contained: no dependencies beyond httpx and the MCP SDK.

Requires a Xen Orchestra instance (XO from sources or the Xen Orchestra Appliance, 5.x with /rest/v0). XO is the management plane this tool talks to — direct per-host XAPI access is out of scope for v0.1. Do NOT use for Proxmox VE — use proxmox-aiops.

What works

• CLI (xcpng-aiops ...): init, overview, vm list/get/stats/health-rca/start/stop/reboot/migrate, host list/get/missing-patches, pool list/get/posture, sr list/get/vdis/usage-rca/rescan, snapshot list/create/delete/revert, backup jobs/logs/failure-rca, task list, secret set/list/rm/migrate/rotate-password, doctor, mcp.
• MCP server (xcpng-aiops mcp or xcpng-aiops-mcp): 29 tools (19 read, 8 write, 2 undo), every one wrapped with the bundled @governed_tool harness.
• Four flagship RCA analyses (cause + action structured output): VM health, SR usage, backup-job failures, pool patch & HA posture.
• Encrypted credentials: the XO authentication token lives in an encrypted store ~/.xcpng-aiops/secrets.enc (Fernet + scrypt) — never plaintext on disk. Unlock with a master password from XCPNG_AIOPS_MASTER_PASSWORD (MCP/CI) or an interactive prompt (CLI).
• Reversibility: vm_start ↔ vm_stop record each other as inverses; vm_migrate captures the REAL source host before moving and records "migrate back"; snapshot_create captures the created snapshot's REAL id from the XO response and records "delete THAT snapshot". Irreversible ops (snapshot_delete, snapshot_revert, vm_reboot) capture prior state for the audit record and honestly declare no undo.
• Safety: destructive CLI ops require double confirmation and support --dry-run; every write MCP tool takes a dry_run preview (no write call, no undo recorded).
• Self-lockout guard (partial — read this): Xen Orchestra is commonly a VM on a pool it manages, and stopping that VM kills the API this tool talks to — vm_start can then no longer be sent, so recovery needs hypervisor console access (xe vm-start). Set xo_self_vm_uuid on the target (xcpng-aiops init asks) and vm_stop refuses exactly that uuid — on --dry-run as well, since a preview that green-lights a call the tool will then refuse is reporting the wrong outcome. If you do not set it there is no protection at all: XO's REST API exposes no self endpoint and its token carries no claims, so the tool cannot discover which VM it runs on, and it fails open rather than guess. The dry_run preview adds a weaker selfVmHint when a VM's reported IP matches the configured XO host — that is a coincidence worth checking, not a finding, and it never blocks (it sees nothing without the guest agent and fires on every VM behind a shared proxy).

Capability matrix (29 MCP tools)

Flagship RCAs

1. vm_health_rca — VMs halted unexpectedly (auto-poweron / HA restart priority set), paused/suspended VMs, running VMs without guest tools, CPU/memory pressure from RRD stats → cause + action per finding.
2. sr_usage_rca — SRs ranked by physical fullness (near-full ≥ 85%, critical ≥ 95%), thin-provision overcommit (virtual allocation > capacity), orphaned VDIs (attached to no VM) with reclaimable bytes per SR.
3. backup_failure_rca — failed/skipped/interrupted XO backup runs classified: vdi-chain (coalesce not finished), quiesce (guest VSS), transport (remote unreachable), storage-full, unknown — with per-job counts and sample messages.
4. pool_patch_ha_posture — hosts missing patches, hosts pending reboot, version skew across a pool's hosts (breaks live migration / rolling updates), multi-host pools without HA.

What this tool does, and does not, decide

It delivers XCP-ng operations — reads and writes — accurately and efficiently, and records every one of them. It does not decide whether a write is allowed to happen. That is the agent's judgement, or the permission of the Xen Orchestra account whose token you connect it with: give that XO user a read-only ACL, or scope its token down, and the writes fail at Xen Orchestra — the place that actually owns the permission.

So there is no read-only switch, no policy file, no approval gate to configure. The one thing the tool guarantees is that nothing is silent: every call, over MCP and over the CLI alike, lands an audit row in ~/.xcpng-aiops/audit.db, and reversible writes still capture their before-state and record an inverse.

Each tool declares a risk_level, kept in agreement with its [READ]/[WRITE] documentation tag by a test, and carried into the audit row as a descriptive tier — so a reviewer can see at a glance that a row was a high-risk snapshot delete. It is a label, not a gate.

Running a smaller / local model? See agent-guardrails.md — it lists the guardrails this tool now enforces for you (so you don't spend prompt budget restating them) and gives a ready-made system prompt for what's left.

Quick start

As a Claude Code plugin

One install gives an agent both the skill and the MCP server:

/plugin marketplace add AIops-tools/marketplace
/plugin install xcpng-aiops@aiops-tools

The MCP server is fetched with uv and pinned to the package version this plugin declares, so an audit row can be traced back to the code that wrote it. Credentials are still configured with xcpng-aiops init — see below.

As an OpenClaw plugin

The same bundle is published on ClawHub, where one install delivers the skill and its MCP server together:

openclaw plugins install clawhub:@zw008/xcpng-aiops
openclaw skills info xcpng-aiops          # expect: Visible to model: yes

Restart the OpenClaw gateway afterwards so it loads the plugin. The MCP server is fetched with uv, pinned to this exact release, so uvx has to be on PATH — without it the skill still installs but reports Visible to model: no. Credentials are configured exactly as below.

As a CLI or standalone MCP server

uv tool install xcpng-aiops
xcpng-aiops init        # interactive wizard: XO URL + encrypted token
xcpng-aiops doctor      # verify config, encrypted store, XO reachability + pool count
xcpng-aiops overview    # one-shot fleet health summary

init writes ~/.xcpng-aiops/config.yaml (non-secret connection details) and stores the XO token encrypted in ~/.xcpng-aiops/secrets.enc. Example config it produces:

targets:
  - name: xo1
    url: https://xo.example.com   # the XO web origin (management plane)
    verify_ssl: true              # set false only for self-signed lab certs
    api_path: /rest/v0

Create the token in the XO UI (user menu → Personal tokens) or with xo-cli --createToken. For non-interactive use (MCP server, CI, cron) export the master password so the store can be unlocked without a prompt:

export XCPNG_AIOPS_MASTER_PASSWORD='your-master-password'

Where that password then lives: an exported variable is readable by every process this shell starts and is recorded by shell history. On a shared or long-lived host, prefer the interactive prompt, or inject it from a secret manager for the life of the one command that needs it.

MCP client config

{
  "mcpServers": {
    "xcpng-aiops": {
      "command": "uvx",
      "args": ["--from", "xcpng-aiops", "xcpng-aiops-mcp"],
      "env": { "XCPNG_AIOPS_MASTER_PASSWORD": "your-master-password" }
    }
  }
}

Env-block caveat: MCP clients launch the server with a minimal environment — your shell profile's exports are not inherited. Put XCPNG_AIOPS_MASTER_PASSWORD (and, if you use them, XCPNG_AIOPS_HOME / XCPNG_AIOPS_CONFIG / XCPNG_AUDIT_APPROVED_BY) in the env block above, or the encrypted store cannot be unlocked and every tool returns a teaching error.

Managing secrets