Plesk now ships a Model Context Protocol server as a free extension, which means an AI agent can provision a subscription, diagnose a bounced email, or hunt down the cause of a 500 error by calling the same REST and WP Toolkit APIs a human admin would use — through a chat window instead of a terminal. For hosts already scripting against Plesk's REST API, this is the same surface with a conversational front end and a permission model bolted on top.
It's early. The extension is beta, unsupported, and has known bugs. This guide covers what it does, how to install and connect it, how token scoping works, and where it currently breaks — so you can decide whether it belongs on a production box yet.
What the MCP Server extension does
MCP — Model Context Protocol — is the emerging standard for connecting LLM clients to external tools; Anthropic describes it as a USB-C port for AI applications. Plesk's implementation wraps the existing REST API and WP Toolkit REST API as MCP tools, plus two capabilities the REST API doesn't expose directly: running arbitrary shell commands on the server and uploading files. Connect Claude Desktop, ChatGPT, or Gemini to it and you can ask it to spin up a subscription, pull a domain's Apache or PHP-FPM error log, restart a service, or hand a WordPress bug straight to a coding agent that can also see the site's files.
Requirements
- Plesk Obsidian 18.0.72 or later, Linux only — there's no Windows build.
- A valid SSL certificate on the panel. Self-signed certificates fail the OAuth handshake the extension uses; if you're still on a self-signed cert, renew via Let's Encrypt before you install this.
- Node.js on the machine running the AI client, if you're configuring Claude Desktop with a free-tier subscription.
- Administrator credentials. As of this writing, MCP access for reseller or customer accounts isn't supported — only the panel admin can authenticate, even though the underlying REST API itself has supported reseller and customer logins since 18.0.79.
Step 1 — Install the extension
From the CLI:
plesk bin extension -i mcp
Or through the UI: Extensions → Extensions Catalog → search "MCP" → Install. Installation takes a few minutes and adds an MCP Server panel under Extensions once it finishes.
Step 2 — Authenticate and grab the endpoint
Open the MCP Server extension in Plesk and start the OAuth flow — it issues a client the way any
OAuth-based integration would. The extension exposes a discovery endpoint at
https://your-panel-hostname:8443/.well-known/oauth-authorization-server that AI clients use to
find the token and authorize URLs automatically.
Step 3 — Point an AI client at it
Claude Desktop, Claude Code, ChatGPT, and Gemini all support MCP as a remote server connection. The exact config differs per client, but the shape is the same — you're registering an HTTP MCP endpoint and letting the client drive the OAuth flow on first connect. For Claude Desktop, add an entry to its MCP server config pointing at your panel's MCP URL and let it prompt you through authorization on first use.
Once connected, ask it something concrete rather than open-ended — "list subscriptions using more than 80% of their disk quota" gets a useful answer; "manage my server" does not.
Step 4 — Scope access with tokens
Don't hand an AI agent full admin access by default. Two controls matter:
- Read-only keys. Restrict an MCP key to read-only at creation, and every write operation — creating a subscription, editing DNS, running a shell command — is rejected at the key level before it reaches the API. Start every new integration here and widen access only once you trust the specific workflow.
- Subscription-scoped tokens. Through the WebPros Dashboard (the multi-server management layer Plesk sits under), you can issue a token scoped to a single subscription rather than the whole server. That's the shape you want if you're ever letting a customer's own AI agent touch their site — it can't see or modify anything outside that one subscription.
Every MCP tool call is written to an activity log visible to the admin, so a post-hoc audit of "what did the agent actually do" doesn't require correlating shell history with panel logs by hand.
What it can and can't do yet
It can: create and manage subscriptions, inspect and restart services, search domain logs, manage WP Toolkit installs (updates, staging, cloning), run shell commands, and upload or edit files on the server — effectively anything the REST API and WP Toolkit API cover.
It can't yet: maintain conversation history across sessions, take fully autonomous action without
a human approving each tool call in the client UI, or authenticate as anyone but the server
administrator. A few operations that exist as plesk bin CLI commands — mailbox creation is the
one users hit first — aren't wrapped as explicit MCP tools, so an agent has to improvise through
the generic shell-command tool, which is less reliable than a purpose-built API call.
Should you turn this on right now
If you run one Plesk server and want to try natural-language troubleshooting, install it on a non-production box first. If you're a host considering exposing this to customers' own AI agents, wait for the extension to leave beta — admin-only auth and no per-tenant support today make it a non-starter for that use case regardless of how the token scoping eventually shapes up. Either way, this is a free extension; it doesn't change Plesk license pricing at any tier, and a licensed Plesk instance is the only prerequisite.