easypanel-mcp-server

MCP Server for full Easypanel control via Claude Code, Cursor and Claude Desktop.

What is this

easypanel-mcp-server connects Claude Code, Cursor and Claude Desktop directly to your Easypanel instance — a modern Docker-based server control panel — through the Model Context Protocol.

Instead of switching between your editor and the Easypanel dashboard, you control everything from inside Claude: deploy from GitHub, update env vars, read live logs, exec into containers, manage domains, databases, volumes and ports, set resource limits, run Docker maintenance and monitor your server — all in natural language.

It maps the Easypanel API to 57 typed tools across 15 categories, plus a single trpc_raw escape hatch that reaches any of Easypanel's ~350 procedures (40+ namespaces) for everything not covered by a dedicated tool. It speaks both API generations — the tRPC API of panels ≤ 2.30 and the new RPC layer introduced in Easypanel 2.31 — auto-detecting which one your panel uses. Every destructive action is gated behind an explicit confirmation, every response opens with a context banner so Claude always knows what it is touching, and an optional read-only mode lets you connect safely to a production panel.

📖 API reference: docs/easypanel-api.md — architecture, the 43 namespaces, confirmed procedures mapped tool-by-tool, and how to discover new ones.

Compatibility

Easypanel 2.31 replaced its internal tRPC API with a new RPC layer (/api/rpc/*, OpenAPI at /api/openapi.json). On panels ≥ 2.31, every v1.x call that carries parameters fails with 400 Input validation failed.

v2.x detects the generation with a single probe request on first call (cached). To skip detection, set EASYPANEL_API_FLAVOR=trpc (≤ 2.30) or EASYPANEL_API_FLAVOR=rpc (≥ 2.31).

Prerequisites

• Easypanel instance running and accessible
• API token — generate at Easypanel → Settings → API → Generate Token
• Node.js ≥ 18 and Claude Code, Cursor or Claude Desktop

Quick start

Option A — npx (no install needed)

Add .mcp.json to your project root:

{
  "mcpServers": {
    "easypanel-mcp": {
      "command": "npx",
      "args": ["-y", "easypanel-mcp-server"],
      "env": {
        "EASYPANEL_URL": "https://your-panel.example.com",
        "EASYPANEL_TOKEN": "your-api-token"
      }
    }
  }
}

Option B — local build

git clone https://github.com/helbertparanhos/easypanel-mcp-server
cd easypanel-mcp-server
npm install && npm run build

{
  "mcpServers": {
    "easypanel-mcp": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/easypanel-mcp-server/dist/index.js"],
      "env": {
        "EASYPANEL_URL": "https://your-panel.example.com",
        "EASYPANEL_TOKEN": "your-api-token"
      }
    }
  }
}

Cursor — reuse env vars across projects

In Cursor Settings → Tools & MCPs → Environment Variables, set:

• EASYPANEL_URL = https://your-panel.example.com
• EASYPANEL_TOKEN = your-api-token

Then your .cursor/mcp.json uses references that apply automatically to every project:

{
  "mcpServers": {
    "easypanel-mcp": {
      "command": "npx",
      "args": ["-y", "easypanel-mcp-server"],
      "env": {
        "EASYPANEL_URL": "${EASYPANEL_URL}",
        "EASYPANEL_TOKEN": "${EASYPANEL_TOKEN}"
      }
    }
  }
}

Environment variables

Adding context to a project

Place this in your project's CLAUDE.md so Claude knows which Easypanel project and service it should operate on by default:

## Easypanel
Project: `my-project` | Service: `my-api` | Branch: `main`
Repo: `owner/repo`

No folder copying needed — one MCP install serves all your projects.

Use cases

"Deploy my app" — Claude lists projects, inspects the current service, triggers deploy_service, then watches list_actions until it completes.

"Why is my service down?" — Claude calls get_service_error, get_service_logs and get_build_logs, and can exec_in_container to inspect files/env live.

"Add DATABASE_URL to staging" — Claude reads current env vars with get_env_vars, adds only the new key with set_env_var (never wipes others), and reminds you to redeploy.

"Give this service 512MB and half a core" — Claude calls set_service_resources (reads current limits and merges your change) and reminds you to restart.

"Persist /app/data and expose port 5432" — Claude calls create_mount (named volume) and create_port, both applied on the next deploy.

"My disk is full" — Claude runs get_storage_stats, then cleanup_docker_images or prune_docker (with confirmation) to reclaim space.

"Show me the Traefik dashboard config" — for anything without a dedicated tool, Claude uses trpc_raw to call the procedure directly.

Available tools (57)

⚠️ = requires confirm: "CONFIRMO". For exec_in_container, create_mount, create_port and trpc_raw the confirmation is conditional (only for destructive commands, sensitive host-path bind mounts, privileged ports < 1024, and write mutations respectively).

Full tool descriptions with parameters are in llms.txt. For the underlying API (both generations), see docs/easypanel-api.md.

trpc_raw — reach any of the ~350 procedures

Covering every Easypanel procedure with a typed tool isn't practical, so anything without a dedicated tool is reachable directly:

// read (default)
{ "procedure": "certificates.listCertificates" }
{ "procedure": "traefik.getDashboard" }

// write — requires isMutation:true AND confirm:"CONFIRMO"
{ "procedure": "branding.updateSettings", "input": { /* ... */ },
  "isMutation": true, "confirm": "CONFIRMO" }

Useful namespaces only reachable via trpc_raw: traefik.*, branding.*, cloudflareTunnel.*, box.*, middlewares.*, notifications.*, volumeBackups.*, databaseBackups.*, wordpress.*, git.*, update.*.

Procedure names use dot notation on both API generations — the client translates to the right transport. On panels ≥ 2.31 the client also validates the procedure against the panel's own OpenAPI spec, fail-closed: a trpc_raw read only executes if the spec documents the procedure as a query, so writes can't sneak past the readonly mode or the confirmation gate.

Safety features

Context banner

Every response that touches a specific project/service starts with:

[Contexto ativo: projeto="my-project" | serviço="my-api"]

Claude always knows what it is modifying before taking any action.

Confirmation guard

Destructive or production-impacting actions return BLOQUEADO until they receive confirm: "CONFIRMO":

{
  "status": "BLOQUEADO",
  "acao": "stop_service",
  "alvo": "serviço \"api\" (usuários perderão acesso)",
  "instrucao": "Para confirmar, passe o parâmetro: confirm: \"CONFIRMO\"",
  "aviso": "⚠️  Esta ação pode ser IRREVERSÍVEL. Confirme apenas se tiver certeza."
}

This gates project/service deletion, stop/rename, env/domain removal, database destruction, the global server ops (prune_docker, restart_panel, reboot_server), and — conditionally — dangerous container commands, sensitive bind mounts, privileged ports and raw mutations.

Read-only mode

Set MCP_ACCESS_MODE=readonly to block every write at the source (client.mutate), covering both curated tools and trpc_raw. Reads remain available — perfect for a production panel you only want to inspect.

Raw escape-hatch controls

trpc_raw validates the procedure name (namespace.procedure, no path/query injection), requires the input to be an object (≤ 50KB), and demands CONFIRMO for any mutation. Set EASYPANEL_RAW_DISABLED=1 to turn it off entirely.

Secret redaction

list_users strips apiToken, twoFactorSecret and password fields before returning — only id, email, admin, twoFactorEnabled and createdAt reach the model.

Safe env vars (read-modify-write)

set_env_var and delete_env_var read the current state, apply only the requested change, and write back. The Easypanel API replaces the entire env string on every update — without this protection it is easy to accidentally wipe all variables at once.

Sensitive value masking

get_env_vars masks values whose key matches *SECRET*, *PASSWORD*, *TOKEN*, *KEY* by default. Pass include_values: true to reveal.

Token never leaks

HTTP errors and WebSocket failures are logged to stderr and surfaced to the model as a generic message — the bearer token (sent in the WebSocket query string, as Easypanel requires) never reaches the model context.

Input validation

projectName / serviceName are validated against ^[a-z0-9][a-z0-9_-]*$ before being used to build a Docker service name or WebSocket query (defense-in-depth against target confusion / parameter injection). Ports are validated as integers 1–65535; resource values must be positive numbers.

Companion skill /ep

Install the workflow skill for guided deploy operations in Claude Code:

mkdir -p ~/.claude/skills/ep
cp skill/SKILL.md ~/.claude/skills/ep/SKILL.md

Then use /ep for an interactive deploy workflow without needing to remember tool names.

How it works

The Easypanel panel talks to its backend over tRPC (/api/trpc/<router>.<procedure>), not a public REST API. This server uses the same endpoints:

• Reads are tRPC queries; writes are tRPC mutations — see docs/easypanel-api.md.
• Live logs, container exec and Docker events use the panel's WebSocket channels (/ws/serviceLogs, /ws/containerShell, /ws/dockerEvents) — the same ones the UI uses — so they work without the licensed Advanced Logs (Loki).
• A few input schemas (mounts, ports, resources) were validated against a live Easypanel and are documented in the API reference.

Known limitations

• WordPress / Box service types — not exposed as dedicated tools; reach them via trpc_raw (e.g. wordpress.inspectService, box.createService).
• trpc_raw reads bypass read-only mode — read-only blocks writes only. A raw read can return sensitive data; use EASYPANEL_RAW_DISABLED=1 in untrusted environments.
• Cluster tools — list_nodes returns the local node only on single-server setups (no Swarm cluster).
• Docker events are real-time only (no history) — an idle server may return an empty window.

Testing without Claude

npx @modelcontextprotocol/inspector dist/index.js

Opens a browser UI where you can call any tool manually and inspect the response.

Comparison with similar packages

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md for how to add tools, report bugs and open PRs.

• Bug report
• Feature request

👤 Author

Created by Helbert Paranhos from Strat Academy.

If this project was useful, consider giving it a ⭐ and following Strat Academy for more AI automation content.

📄 License

MIT © Helbert Paranhos / Strat Academy

See LICENSE for details.