Wakeonlan Mcp Server

Overview

Wake-on-LAN by name for the machines on your network. Wake a sleeping desktop, GPU box, NAS, or lab machine, wait until it answers on a port like SSH, check whether a host is up without waking it, and troubleshoot a wake that didn't take. Runs on any machine on the same LAN as its hosts.

Tools

Skills

Served over MCP as skill://wakeonlan-guide/SKILL.md and shipped in the npm package under skills/.

Capability reference

wol_wake_host _tool

• alias (from wol_list_hosts, case-insensitive) plus optional wait_for_s, 0–55 seconds, default 30; 0 sends and returns without checking. Sends 3 packets 500 ms apart, probing the check port once before sending and every 2 s after the first packet
• state is already_awake, awake (with time_to_answer_ms, an upper bound at the 2 s poll interval), not_reachable, or unverified (unverified_reason: wait_disabled or no_address); the last two carry a guidance next step
• Fails as unknown_host, off_segment (nothing sent: this machine has no interface on the host's subnet), retryable wake_in_progress (nothing sent: another call is already waking that host, and one wake per host runs at a time), or retryable socket_error, whose data carries the failed stage and packets_sent

wol_check_host _tool

• alias only: one TCP connect to the profile's address and check_port (22 unless the profile sets another), 1.5 s timeout, no magic packet. The port is not an input, so the tool can't scan arbitrary ports
• outcome is open, refused (the machine answered, nothing listens on that port), or no_answer; reachable is true only for open, and latency_ms is present for open and refused
• Fails as unknown_host, or no_address when the profile has no address to probe

wol_list_hosts _tool

• No input. Returns every profile in config order: alias, description, mac, address, check_port, wol_port, broadcast with broadcast_source (configured, derived, or unresolved), and secureon_set (never the password)
• on_segment, with the sending interface and local_address when true, is resolved against this machine's interfaces on every call; nothing is sent or probed, so it says nothing about whether a host is up
• config_source (file, inline, or none) and config_path name where the profiles came from

wol_list_reference _tool

• topic: packet-format, prerequisites, sleep-states, troubleshooting, host-profiles, or sender-environment
• Static markdown with no network access; every response lists all topics for navigation

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Wake-on-LAN-specific:

• Operator-configured targets only: callers pass an alias, never a MAC, IP, broadcast address, or port, so an agent or HTTP caller can wake and probe only the hosts you listed. SecureOn passwords stay in the profile, never returned or logged
• Subnet-directed broadcasts: the broadcast address is derived from a host's IPv4 address and the matching local interface when the profile omits it, and a host on no local subnet is refused before anything is sent instead of being routed away silently
• Wakes are confirmed by a TCP connect to a per-host check_port (SSH by default; RDP or SMB for Windows), not ICMP
• The LAN layer is Node's own node:dgram, node:net, and node:os, with no third-party networking dependency

Agent-friendly output:

• Results, not errors, for every wait outcome: state plus a guidance next step naming the tool or reference topic to call
• Typed failures with recovery hints: unknown_host lists up to 20 configured aliases, off_segment names the local subnets it compared, and on macOS socket_error points at the Local Network permission
• refused vs no_answer: a machine that is on but not listening on its check port reads differently from one that never answered. A wake still ends not_reachable either way; probe.last_outcome and guidance say which, so a wrong check_port can be told apart from a failed wake

Getting started

Add the following to your MCP client configuration file, pointing WOL_HOSTS_FILE at your hosts file.

{  "mcpServers": {    "wakeonlan-mcp-server": {      "type": "stdio",      "command": "bunx",      "args": ["@cyanheads/wakeonlan-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info",        "WOL_HOSTS_FILE": "~/.config/wakeonlan/hosts.json"      }    }  }}

Or with npx (no Bun required):

{  "mcpServers": {    "wakeonlan-mcp-server": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@cyanheads/wakeonlan-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info",        "WOL_HOSTS_FILE": "~/.config/wakeonlan/hosts.json"      }    }  }}

The Claude Desktop .mcpb bundle (the install badge above) asks for a hosts file or inline hosts JSON when you install it.

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 WOL_HOSTS_FILE=~/.config/wakeonlan/hosts.json bun run start:http# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• A machine attached to the same LAN segment as the hosts it wakes. Run the server on that machine's OS, or in Docker with host networking on Linux: a container on a default bridge network, Docker Desktop on macOS or Windows, or WSL2 in its default NAT mode can't put a broadcast on the LAN.
• Targets with Wake-on-LAN enabled in firmware and armed by the OS. wol_list_reference with topic prerequisites has the Windows, Linux, and macOS settings.

macOS: Local Network permission

On macOS 15 and later, sending a UDP broadcast or connecting to a LAN address needs Local Network access, and macOS grants it to the app that launched the server rather than to node:

Grant or check it under System Settings > Privacy & Security > Local Network. The first send can fail while the alert is pending, so wol_wake_host may return socket_error; retry after allowing. If that list shows a node entry rather than your client app, enable the node entry. A denied permission also makes wol_check_host read no_answer. Windows and Linux have no per-app gate. wol_list_reference with topic sender-environment covers the rest, including the macOS 15.5+ subnet exemption.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/wakeonlan-mcp-server.git

2. Navigate into the directory:

cd wakeonlan-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env# edit .env and set WOL_HOSTS_FILE or WOL_HOSTS

Configuration

See .env.example for the full list of optional overrides.

Host profiles

Profiles are a JSON array, read once at startup from WOL_HOSTS_FILE or WOL_HOSTS. Set one of the two: both set is a startup error, and neither set starts the server with no hosts, which wol_list_hosts explains. WOL_HOSTS_FILE must be absolute (after ~/ expansion), because a stdio server runs in the MCP client's working directory, and must name a regular file of at most 1 MiB: a directory, a pipe, or a device such as /dev/stdin is a startup error. Editing the profiles takes a restart.

[  {    "alias": "gpu-box",    "description": "Desktop with the training GPU; SSH on 22.",    "mac": "00:00:5e:00:53:01",    "address": "192.0.2.50"  },  {    "alias": "nas",    "mac": "00-00-5E-00-53-02",    "address": "nas.home.arpa",    "broadcast": "192.0.2.255",    "check_port": 445,    "secureon": "00:00:5e:00:53:ff"  }]

An empty string leaves an optional text field (address, broadcast, secureon, description) unset. The server refuses to start, naming the entry and field, on an unknown key, an invalid value, a duplicate alias, or a profile with neither broadcast nor an IPv4 address. Whether this machine sits on a host's subnet is checked per call, not at startup, since interfaces change. wol_list_reference with topic host-profiles has an example per OS.

HTTP exposure

A startup guard refuses any HTTP deployment that would let an unauthenticated caller reach the tools from beyond this machine, or from a web page through your browser. Stdio is unaffected.

DEV_MCP_AUTH_BYPASS counts as unauthenticated. With MCP_ALLOWED_ORIGINS unset, requests from non-loopback browser origins are rejected, so a web page can't drive a loopback endpoint through DNS rebinding; * turns that check off, which is why it needs auth. Host profiles are server-wide: every authenticated caller can wake the same hosts.

Running the server

Local development

• Build and run the production version:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:http# orbun run start:stdio

• Run checks and tests:

  sh

  bun run devcheck  # Lints, formats, type-checks, and morebun run test      # Runs the test suite

Docker

The image is Linux-only. It can wake hosts only when run with --network host (or on a macvlan network) on a Linux machine attached to their LAN, such as a Raspberry Pi, NAS, or home server that already runs Docker. On a default bridge network the container sees only Docker's private subnet, so wol_wake_host fails with off_segment before sending anything. Docker Desktop on macOS and Windows runs containers in a VM, so its broadcasts can't reach the LAN in any network mode.

Pull the published multi-arch image (linux/amd64, linux/arm64):

docker pull ghcr.io/cyanheads/wakeonlan-mcp-server:latest

Or build it from a clone of this repository with docker build -t ghcr.io/cyanheads/wakeonlan-mcp-server:latest .. Then add it to your MCP client configuration on that machine. The hosts file is mounted read-only from an absolute host path, and WOL_HOSTS_FILE names where it sits inside the container:

{  "mcpServers": {    "wakeonlan-mcp-server": {      "type": "stdio",      "command": "docker",      "args": [        "run", "-i", "--rm",        "--network", "host",        "-v", "/path/to/hosts.json:/etc/wakeonlan/hosts.json:ro",        "-e", "MCP_TRANSPORT_TYPE=stdio",        "-e", "WOL_HOSTS_FILE=/etc/wakeonlan/hosts.json",        "ghcr.io/cyanheads/wakeonlan-mcp-server:latest"      ]    }  }}

The container runs as the image's bun user (uid 1000), which must be able to read the hosts file. Without MCP_TRANSPORT_TYPE=stdio the image serves Streamable HTTP on port 3010, bound to loopback, which under host networking is the host's own; any other bind needs MCP_AUTH_MODE jwt or oauth (see HTTP exposure). Logs go to /var/log/wakeonlan-mcp-server. OpenTelemetry peer dependencies are installed by default; build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging; nothing persists, so ctx.state goes unused
• Register new tools in the createApp() arrays in src/index.ts
• Every OS boundary (sockets, interfaces, clock, platform, filesystem) is an injected seam: tests never open a real socket, and live checks run on loopback only

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.