pbs-mcp

MCP server for Proxmox Backup Server. Exposes datastore status, snapshot inventory, garbage collection, verify, and prune over the PBS REST API as 17 LLM-callable tools. Designed for the Model Context Protocol.

Türkçe için → README.tr.md

Why

PBS already has a polished web UI. This server is for the cases where the UI isn't where you are — answering "is anything broken?" from a chat assistant, or wiring PBS state into a homelab agent that schedules verifies and prunes based on real conditions.

Related MCP server: duplicati-mcp

Tools (17)

Start with pbs_health_overview for any "is PBS okay?" question — it fetches datastore status, GC stats, groups, snapshots, node status, and recent tasks concurrently and condenses them into one verdict, replacing five separate tool calls.

Write tools require both PBS_ALLOW_WRITE=true in the environment and confirm=true in the call itself. Restore is intentionally out of scope — the standard proxmox-mcp already handles restore from PBS via archive= on the Proxmox VE side.

Setup

1. Create an API token in PBS

In a shell on the PBS host (pct enter 205 from Proxmox if PBS lives in an LXC, otherwise just SSH):

# Generate a token under root@pam
proxmox-backup-manager user generate-token root@pam mcp

# Grant it admin on your datastore
proxmox-backup-manager acl update /datastore/<your-datastore> \
  DatastoreAdmin --auth-id 'root@pam!mcp'

The generate-token output includes a value field — that's the secret, shown only once. Save it.

Why DatastoreAdmin? PBS performs an owner check on prune. Either you give the token DatastoreAdmin (this), or you keep moving backup ownership with change-owner after every push. Admin scope is simpler and stays inside one datastore.

2. Configure the MCP server

git clone https://github.com/ahmetem/pbs-mcp.git
cd pbs-mcp
cp .env.example .env
# Edit .env: fill in PBS_HOST, PBS_TOKEN_ID, PBS_TOKEN_SECRET
pip install -e .

3. Register with your MCP client

For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "pbs": {
      "command": "python",
      "args": ["/absolute/path/to/pbs-mcp/pbs_mcp.py"]
    }
  }
}

Restart the client. The pbs_* tools should appear.

Safety model

• Read-only by default. Out of the box PBS_ALLOW_WRITE=false, so every state-changing tool refuses regardless of confirm.

• Two-key requirement on writes. PBS_ALLOW_WRITE=true opens the door; each call still needs confirm=true. Two independent toggles, two intentional actions.

• Async tasks return UPIDs, not results. run_gc and run_verify kick off work and hand back a UPID immediately. Poll with get_task_status. This prevents the MCP request from blocking for hours.

• Token scoped to one datastore. ACLs live at /datastore/<name>, not at /. A leaked token can't read PBS user lists or remote sync configs.

Notes / gotchas

• First-call cache lag: PBS caches ACLs for a few seconds. GET requests retry once automatically after a short delay on 403 / connect errors, so a freshly granted token usually just works. If it still fails, wait a few seconds and retry.

• Token vs user permissions: PBS API tokens get the intersection of the parent user's ACLs and the token's ACLs. With root@pam!mcp the parent is unrestricted, so only the token's ACL matters in practice.

• Self-signed cert: PBS ships with a self-signed cert. The default PBS_VERIFY_TLS=false is fine for a LAN setup. For real CAs, set PBS_VERIFY_TLS=true and PBS_CA_BUNDLE to a PEM file.

• UPIDs are tied to their creator: a UPID created by a now-deleted user becomes unreadable. Don't recycle PBS users while there are pending tasks.

Development

pip install -e .[dev]
pytest

Tests stub the HTTP layer — no PBS instance needed.

License

GPL-3.0-or-later. See LICENSE.