Proxmox MCP

Proxmox + AI - for Claude, the ChatGPT desktop app, GitHub Copilot, Gemini CLI, and any agent that speaks MCP

Unofficial. Community-built Claude Code Skill and MCP server for the Proxmox API. Not affiliated with, endorsed by, or sponsored by Proxmox Server Solutions GmbH.

[Proxmox demo - animated preview]

_{▶ Watch the 30-second demo - demo data is simulated; every command shown exists in the real CLI.}

Also from Servosity. Backup & DR is Servosity's own field - the first-party Servosity connector brings this same fleet-wide, local-mirror approach (fleet attention, stale backups, restores, QBR reporting) to Servosity Backup and DR.

Comparing backup vendors? One of our MSP partners told us, "My guys now spend a tenth of the time on backups." Fixing is included in every Servosity co-managed backup plan, and our engineers tell you what they did. See how co-managed backup works

Read-only Proxmox VE and Backup Server CLI that tells you, across every client cluster, which guests have a good, verified, recent restore point and which do not. Works with the AI you already use - ChatGPT (desktop app), Claude Desktop, Codex, Claude Code, Claude Cowork, and GitHub Copilot - plus Gemini CLI. Free, open source, runs on your laptop. Built for MSP owners. No code required.

Works with your agent

Security note (until this connector is reprinted). Do not run HTTP mode (--transport http), not even on localhost. This build checks a token, but it predates the fix release's request size and time limits. The MCP tools can also read your data with your credentials. Use the default stdio transport. If you set up a tunnel for a remote agent, shut it down. A fix is on the way.

The six agents MSP owners actually use (self-serve, works today):

The ChatGPT desktop app runs proxmox-mcp locally and shares MCP configuration with Codex (OpenAI). ChatGPT on the web needs a remote (HTTP) server, and this connector's HTTP mode is paused until it is reprinted (see the security note in mcp-install.md), so msp-skills does not support it yet. If your team works there, a hosted gateway is the better pick; check that your AI app is on the gateway's setup list.

Also for the Microsoft and Google stacks

Big install base, but an honest heads-up: Gemini CLI runs the local binary you just installed, while Microsoft 365 Copilot and the Gemini app need a remote server, which msp-skills does not support yet.

Skill-native agents (also covered): Hermes and OpenClaw read this skill's SKILL.md directly and speak MCP - see their install sections below. Also works with Cursor, Windsurf, Cline, Continue.dev, and Zed via MCP. Full per-tool wire-up: docs/which-agent.md.

Run more than one agent? Install across all 51+ supported agents in one command: npx skills add Servosity/msp-skills@latest (requires Node.js, then run the per-skill installer for the CLI/MCP binaries). See docs/which-agent.md.

Install in 60 seconds

Fastest for Claude Desktop - one-click .mcpb

Download Proxmox MCP (.mcpb) - then open Claude Desktop > Settings > Extensions and select the file. One click, no JSON, no shell. (Browse every Proxmox release on the releases page.)

Prefer the Claude Code plugin? Add the marketplace once, then install - works immediately, no directory listing required:

/plugin marketplace add Servosity/msp-skills/plugin install proxmox@msp-skills

Path A - paste one prompt into your AI agent (recommended)

Copy this into Claude Code, Codex CLI, or Claude Cowork:

Install the Proxmox Skill and MCP server from Servosity/msp-skills in this agent workspace. If this workspace uses a POSIX shell (macOS, Linux, WSL, or Bash), run bash <(curl -fsSL https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.sh). If it uses Windows PowerShell, run iwr -useb https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.ps1 | iex. Then authenticate per the README and run proxmox-cli --help to explore.

The same prompt works in any agent that can run shell.

Path B - run the installer yourself

Windows (PowerShell):

iwr -useb https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.ps1 | iex

macOS / Linux:

bash <(curl -fsSL https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.sh)

The installer drops both proxmox-cli (the CLI) and proxmox-mcp (the MCP server) into your user bin path. Claude Code, Codex, and Cowork discover the Skill via SKILL.md in this directory.

Verify:

proxmox-cli --version

Upgrade to the latest version

The installer always fetches the current release - re-run it to upgrade:

macOS / Linux:

bash <(curl -fsSL https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.sh)

Windows (PowerShell):

iwr -useb https://raw.githubusercontent.com/Servosity/msp-skills/main/skills/proxmox/install.ps1 | iex

Claude Desktop .mcpb users: download the latest .mcpb (top of this section) and re-select it in Settings > Extensions. Claude Code plugin users: /plugin update proxmox@msp-skills.

Add to Claude Desktop, the ChatGPT desktop app, GitHub Copilot, Gemini CLI, or another MCP client

After the installer runs, see mcp-install.md and docs/which-agent.md for the per-agent wire-up - one section per agent, including the GitHub Copilot servers key and the ChatGPT desktop app. Claude Desktop's Settings > Extensions panel is the simplest path; the MCP config block (for users who prefer editing JSON) is documented in mcp-install.md.

Install for Hermes

From the Hermes CLI:

hermes skills install servosity/msp-skills/skills/proxmox --force

Inside a Hermes chat session:

/skills install servosity/msp-skills/skills/proxmox --force

Hermes speaks MCP natively, so it can also use the proxmox-mcp server directly - same install path, same env vars.

Install for OpenClaw

Tell your OpenClaw agent (copy this):

Install the proxmox skill from https://github.com/servosity/msp-skills/tree/main/skills/proxmox. The skill defines how its required CLI (proxmox-cli) can be installed via the openclaw: frontmatter block.

OpenClaw isn't generally available yet; the frontmatter wiring is pre-shipped and will activate the moment OpenClaw launches.

Authenticate

Create a privilege-separated API token for each product: PBS gets the Audit role on /; PVE gets PVEAuditor on /, plus Datastore.AllocateSpace and VM.Backup if your backups live on PVE storage (see What API token does it need?). Then:

export PROXMOX_HOST=pve1.example.lanexport PROXMOX_API_TOKEN='monitor@pve!agent=<secret>'      # USER@REALM!TOKENID=SECRETexport PROXMOX_TLS_FINGERPRINT='<PVE certificate SHA-256>'  # stock Proxmox certificates are self-signed# Optional, for PBS verify, datastore and sync data:export PROXMOX_PBS_HOST=pbs.example.lanexport PROXMOX_PBS_API_TOKEN='monitor@pbs!agent:<secret>'  # colon before the secretexport PROXMOX_PBS_TLS_FINGERPRINT='<PBS certificate SHA-256>'proxmox-cli doctor

Managing several clients? Add one entry per client cluster instead; clusters.toml stores only the names of the env vars that hold each token:

proxmox-cli clusters add acme --client "Acme Dental" --pve-host pve1.acme.lan --pve-token-env ACME_PVE_TOKEN --pve-fingerprint '<sha256>' --pbs-host pbs.acme.lan --pbs-token-env ACME_PBS_TOKEN --pbs-fingerprint '<sha256>'proxmox-cli clusters list

doctor checks config, paths, API reachability and whether the PVE token authenticates. Run proxmox-cli posture --agent to confirm the token can also see guests, jobs and backups: anything it cannot read is listed under fetch_failures.

What this skill does

Full command reference: guide.md. For the AI-agent operating contract (--agent, --dry-run, when to confirm before mutating), see AGENTS.md.

What makes this different

Most Proxmox MCP servers are built to operate one cluster: list VMs, start and stop them, take snapshots, with the dangerous actions behind a flag. Many MSPs will never hand an AI agent that token. This one sends Proxmox nothing but reads (HTTP GET), so there is no write command to misfire, and with backups on Proxmox Backup Server both tokens can be audit-only.

The answer is the join. "Is this VM backed up?" needs four things Proxmox keeps in different places: whether an enabled job includes it, the newest real archive on any backup storage, whether Proxmox Backup Server verified that archive, and what the last vzdump task log said about that guest. posture does that join for every guest across every client cluster (backup coverage answers the first part: which guests and disks no enabled job includes), failures and rpo cover every guest a job touches, and all of them read live on every run. For inventory questions, sync keeps a local SQLite mirror with full-text search.

The pain this closes

A Proxmox shop with many clients runs one cluster, and often one Backup Server, per client. A green backup task does not mean every guest was backed up: one vzdump job backs up many guests and names the one that failed only inside its task log. Verify failures spread quietly too. In a Proxmox forum thread, a provider whose central PBS takes backups from 100 to 200 servers watched the weekly verify job report a growing number of failed snapshots before the cause was traced. And when an AI-agent management layer for Proxmox was shared on the same forum, administrators pushed back hard on letting an agent near their hosts.

This skill answers the morning questions without asking for that trust:

• posture --only problems - every guest that is uncovered, has no restore point, is stale, failed verification, or whose last attempt failed
• backup coverage --include-disks - every guest or disk no enabled job includes, with the reason
• failures --since 24h - failed and silently incomplete tasks, one row per guest or datastore with the error line
• verify-health - how far a verify failure spread, and the newest verified-good point left
• rpo --days 30 --by-client - per-client proof that every scheduled run produced a restore point

See pain-point.md for the longer narrative.

Frequently asked questions

Does this work with ChatGPT?

Yes, in the ChatGPT desktop app. It runs MCP servers locally (STDIO), the way proxmox-mcp runs, and shares MCP configuration with Codex (OpenAI): add proxmox-mcp under Settings > MCP servers as a STDIO server, or add it once to Codex's ~/.codex/config.toml. ChatGPT on the web is not supported by msp-skills yet: it needs a remote (HTTP) server, and this connector's HTTP mode is paused until it is reprinted; the security note in mcp-install.md explains why. If your team works there, a hosted gateway is the better pick; check that your AI app is on the gateway's setup list.

Does this work with Codex, Cursor, Windsurf, Cline, Copilot, or Gemini?

Yes - all of them speak MCP. Cross-tool install commands are in the matrix above and the deep-dive in docs/which-agent.md.

Do I need to know how to code?

No. The recommended install is to paste one sentence into Claude Code or Codex - your agent reads SKILL.md and does the install. The fallback is a one-line installer per OS (bash or PowerShell). Neither path requires writing code. You'll enter your Proxmox credentials once.

Is my Proxmox data safe?

Your data stays on your machine. The CLI and MCP server are local binaries. The SQLite mirror sits in a directory under your user account. The AI agent only sees what the CLI returns - typically a query result, not raw bulk data. Credentials are read from your environment or your agent's config; never bundled into this repo or transmitted anywhere by MSP Skills.

Can the AI stop, change or delete anything on my Proxmox hosts?

No. The CLI and MCP server only contain read (GET) calls to Proxmox VE and Proxmox Backup Server; there is no command that starts, stops, migrates, restores, prunes or deletes anything. The only writes are local files: the cluster list, a saved token, the local mirror (which the per-endpoint read commands also update, keeping what they return, comments included) and the agent learning store. There is no response cache. Posture, backup coverage, verify-health, failures, capacity, restore-points and rpo always read Proxmox live; a per-endpoint read command answers from the local mirror only when Proxmox cannot be reached, says so on stderr and in meta.source: local, and refuses to with --data-source live. Secret-named fields, such as a third-party storage plugin's password in the storage config, are redacted before anything is cached, stored or shown, and each --cluster keeps its own local store. One side effect to know: reading a storage's status lets Proxmox mount an enabled storage, which its own status daemon already does every 10 seconds.

What API token does it need?

A privilege-separated API token for each product. Proxmox Backup Server: the Audit role on / (read-only). Proxmox VE: PVEAuditor on /. One catch: Proxmox VE lists backup files only to a token that also holds Datastore.AllocateSpace and VM.Backup, so if your backups live on dir, NFS or CIFS storage, or you have not configured the Backup Server here, add a custom role with those two privileges. An audit-only token is reported as "storage not read" and the guest as unproven, never as "no backups". If your backups are on Proxmox Backup Server and PROXMOX_PBS_HOST is set (spelled as in the PVE storage's server field: the same IP or hostname), the CLI reads them from the PBS API and the PVE token can stay audit-only. Without it, a PBS storage is read only through Proxmox VE, which lists it with the PBS credential saved in that storage and so may miss backup groups another credential owns: its restore points count, but a guest without one there is reported unproven, not never, and no guest covered by a backup job on that cluster reads green (a fresh one is archives-incomplete, since a newer backup that failed verification may be hidden). On a stock install the local storage also accepts backups: with an audit-only token it is listed under fetch failures until you remove "VZDump backup file" from its content types or grant it the two privileges. Until then fresh guests keep their verdicts, but a guest with no readable or only an old restore point is reported unproven rather than never or stale (the unread storage could hold one), rpo counts its missed runs as unproven, and posture --fail-on-red exits 1. Grant the roles on / (propagating) and put no NoAccess or narrower ACL below it: Proxmox leaves out of every list the guests and storages a token cannot see, and while the CLI flags a missing or narrowed grant, an explicit NoAccess on one guest is invisible to any read. The PVE token is USER@REALM!TOKENID=SECRET; the PBS token uses a colon before the secret.

Proxmox uses a self-signed certificate. Do I have to turn off certificate checks?

No, and there is no option to. Pin the certificate's SHA-256 fingerprint (shown under the node's Certificates) with PROXMOX_TLS_FINGERPRINT and PROXMOX_PBS_TLS_FINGERPRINT, or per client cluster in clusters.toml. A malformed pin is an error, never silently ignored, and a PVE pin never vouches for the PBS host.

How do I cover many client clusters?

Add one entry per client with proxmox-cli clusters add: its PVE host, optional PBS host and namespace, certificate fingerprints, and the names of the environment variables that hold its tokens (never the tokens themselves). The fleet commands then run across every configured cluster; --cluster or --client narrows to one. For an MCP client, add those token variables to the MCP server's env block too. On a shared Backup Server, give each client cluster its own PBS namespace and declare it as that entry's pbs_namespace (a separate datastore per client is not enough on its own: only a declared namespace scopes a cluster that cannot be read): VMIDs repeat across clusters, so where two clusters write the same namespace the CLI cannot tell whose vm/100 a snapshot is, reports the overlap, and does not count those restore points. A cluster that cannot be read is assumed to write anywhere in its declared pbs_namespace, or, when it declares none, anywhere on its PBS (on any PBS, if its entry names none), so declaring one keeps a client outage from hiding the others' backups.

What does it cost?

Free. Apache-2.0 licensed. You pay only for whichever AI agent you use (Claude, ChatGPT, Codex, etc.), and that's billed by your AI provider, not by us.

Safety model

Every request this skill sends to Proxmox is a read, so the tiers are about what happens on your machine and where output goes.

The strongest control is still the scope you grant the tokens: with backups on Proxmox Backup Server both tokens can be audit-only, so Proxmox itself would refuse a write even if one were ever sent. Backups on PVE storage need two backup privileges before Proxmox VE will list them; governance.md explains the trade-off. Full details, including how to lock it down, are in governance.md.

Status

Beta. Validated against the Proxmox VE and Proxmox Backup Server API surface with mocked responses, and being validated with MSPs running it live against their own clusters in our weekly Build Sessions. RSVP at compoundingteams.com/build-sessions.

Standards. Conforms to the open Agent Skills spec (Anthropic, Dec 2025; 40+ agents). MCP-compatible - works with any MCP-capable agent including Hermes. OpenClaw-ready (frontmatter pre-wired, awaiting OpenClaw launch).

Maintained by Servosity. Apache-2.0 licensed. Built with CLI Printing Press. Last updated: 2026-10-14.