Coolify MCP Server

Manage Coolify from Claude, Cursor, or any MCP client: 44 consolidated tools for deploying, debugging, and operating your self-hosted PaaS in plain English.

📖 coolify-mcp.stumason.dev: what it does, how to install it, and why it is safe to point at production.

This README is the full reference: every tool, every gotcha, every parameter.

Install

You need a running Coolify v4 instance and an API token (Coolify → Settings → API).

Claude Desktop, one-click: download coolify-mcp.mcpb and drag it into Settings → Extensions. You'll be prompted for your Coolify URL and token. No Node install, no JSON editing.

Claude Code:

claude mcp add coolify \
  -e COOLIFY_BASE_URL="https://your-coolify-instance.com" \
  -e COOLIFY_ACCESS_TOKEN="your-api-token" \
  -- npx @masonator/coolify-mcp@latest

Any MCP client (JSON config):

{
  "mcpServers": {
    "coolify": {
      "command": "npx",
      "args": ["-y", "@masonator/coolify-mcp"],
      "env": {
        "COOLIFY_BASE_URL": "https://your-coolify-instance.com",
        "COOLIFY_ACCESS_TOKEN": "your-api-token"
      }
    }
  }
}

Behind Cloudflare Access or an auth proxy? Add --header "Key: Value" args (repeatable). The same config works in Cursor, Claude Code and any other MCP client, and can be repeated for multiple Coolify instances.

Tools

Every tool takes an action parameter; run one with no arguments and it lists what it accepts.

Design

• Token-optimized. Consolidated action-param tools keep the tool list at ~6,600 tokens instead of ~43,000, so the server doesn't eat your context window before you've asked anything.
• Summaries by default. list_* tools return uuid/name/status projections, 90–99% smaller than the raw API measured against a real 21-app estate. get_* tools fetch full detail for one resource.
• Smart lookup. diagnose_app takes a UUID, name, or domain; diagnose_server takes a UUID, name, or IP.
• Actionable responses. Results carry _actions hints (view logs, restart, next page) so the assistant knows the logical next step without extra tokens.
• Verified deploys. deploy with wait: true polls to a terminal status and returns a log tail on failure, instead of "the site returns 200 so it probably worked".

Ask before it hurts

Destructive operations pause and ask you, not the model, on clients that support elicitation: Claude Code and VS Code Copilot today. The prompt states the blast radius before you answer:

EMERGENCY STOP: take down 12 running applications
(api, worker, cockpit, umami, scheduler, mailer, search, billing and 4 more)
across 3 servers?

Confirmation is asked for on stop_all_apps, redeploy_project, restart_project_apps, system disable_api, application / database / service / project / environment deletes, the credential deletes (private_keys, cloud_tokens, github_apps, none recoverable from Coolify once gone), and bulk_env_update across more than three apps. Routine deletes (storages, scheduled tasks, individual env vars, backup schedules) deliberately stay unprompted: a dialog on every delete is how dialogs stop being read. Deleting a resource spells out whether its persistent volumes go with it. delete_volumes defaults to true upstream, so leaving the flag unset is the destructive choice, not the cautious one.

Prompts are skipped where there is nothing to confirm: an emergency stop on an idle estate, or a redeploy of an empty project, just runs.

This is progressive enhancement, not a new requirement: clients without elicitation support (Claude Desktop, claude.ai) behave exactly as before. Once a client does advertise support, a decline, a cancel or a timeout all abort the call.

These tools also carry the MCP destructiveHint annotation, so on a client that honours annotations and supports elicitation you may answer two dialogs in a row: the client's own permission prompt, then this one. That is the client's prompt plus the server's, not a bug. Allowlisting the tool in your client removes the first and leaves this one as the gate.

Set COOLIFY_MCP_ELICITATION=off to turn the confirmations off entirely. It exists for the case where a client advertises elicitation support but does not actually implement it. Without it, every guarded tool would return could not confirm with the user with no way to recover. It is an escape hatch, not a normal setting.

If confirmations time out before you can answer them, raise your client's MCP tool timeout. The prompt runs inside the tool call, and the MCP SDK's default request timeout is 60 seconds. The server aborts cleanly when the client gives up (nothing runs behind your back), but you will see the call fail rather than the dialog you were reading.

Secure by default

Secrets are masked at the API boundary. A client granted "list" access never sees plaintext credentials unless you explicitly opt in with reveal: true:

• env_vars: variable values return as ***
• system list_resources (full mode): webhook HMAC secrets, basic-auth and database passwords, internal/external_db_url connection strings, compose bodies, Traefik labels, nested env vars
• get_database / get_service: the same credential fields are masked on the detail endpoints, and any embedded server row is projected down to uuid/name/ip so its sentinel token and log-drain config never leave the client
• get_server: sentinel and log-drain credentials are always masked, with no reveal
• private_keys: key material is never returned, with no reveal; name, fingerprint and public key identify a key
• deployment get: the raw upstream payload (server settings, log-drain tokens, webhook secrets) never leaves the client; responses are projected

Log output (logs, application_logs, deployment logs) is wrapped in a tamper-evident untrusted-data boundary so a poisoned log line reads as data, not instructions.

Destructive operations also ask a human first; see Ask before it hurts above.

Coolify version compatibility

Works against Coolify v4.0 through v4.2+. Two v4.2 changes are worth knowing about:

• Secrets are hidden by default. From v4.2 Coolify strips sensitive fields from API responses unless the token has sensitive-read scope, so reveal: true can return a variable with no value at all. That is the server withholding it, not a bug here; issue a token with sensitive-read scope if you need plaintext back.
• Member-role tokens are read-only. From v4.2 a token belonging to a Member-role user can view resources but cannot deploy, start, stop, create, update or delete. Those calls return 403. Promote the user or use a token from a role with write access.

State-changing endpoints also moved from GET to POST in v4.2. The client handles this for you across both eras, so no action is needed.

Example prompts

Give me an overview of my infrastructure
Diagnose my stuartmason.co.uk app
Find any issues in my infrastructure
Deploy application {uuid} and wait for it to finish
Update the DATABASE_URL env var for application {uuid}
Create a staging environment in project {uuid}
Restart all applications in project {uuid}
How do I fix a 502 Bad Gateway error in Coolify?

Development

git clone https://github.com/StuMason/coolify-mcp.git
cd coolify-mcp && npm install
npm run build && npm test

COOLIFY_BASE_URL="https://your-coolify.com" COOLIFY_ACCESS_TOKEN="token" node dist/index.js

Evals & red teaming

Because tool descriptions are prompts, evals/ measures whether a model picks the right tool from this surface and whether attacker-controlled tool output can make it misbehave. Deterministic contract snapshots gate every PR; tool-selection and prompt-injection evals run a real model against a mock Coolify backend; a promptfoo red-team battery runs on a schedule. Nothing touches production. See evals/README.md.

Contributions welcome; see CONTRIBUTING.md and the architecture notes in CLAUDE.md.

Work with me

I'm Stu Mason. I build MCP servers, AI integrations and agentic systems for agencies, SMEs and enterprise. This repo is what that work looks like in the open.

• An MCP server for your product. Give Claude, Cursor and every other AI client a proper way into your API, like this one.
• Answers from your own stuff. AI that answers from your documents and data, with the receipts, instead of guessing. Can stay on your own servers.
• Work that runs itself. Jobs on a schedule that sort, check and report, with a person signing off before anything goes out.

White-label under your own name if you're an agency. And if a job doesn't need AI, I'll say so before anyone's paid for anything.

📮 hey@stumason.dev · stumason.dev · coolify-mcp.stumason.dev

Links

• Coolify: the open-source, self-hostable PaaS this server drives
• MCP Registry: listed as io.github.StuMason/coolify
• laravel-coolify: deploy Laravel to Coolify with a dashboard, Artisan commands, and generated Dockerfiles
• Model Context Protocol

MIT © Stu Mason. If this is useful, ⭐ the repo.