rdc — Remote Desktop Control for AI agents

Let a coding agent on your laptop see and operate another computer's desktop over your Tailscale network: take screenshots, move and click the mouse, type, press shortcuts, focus windows, read and write the clipboard. Works with macOS, Linux and Windows targets and plugs into Claude Code (or any MCP client) as a set of tools.

Typical use: your agent is fixing something on a headless Mac mini and hits a dialog that needs a click. Instead of walking over or opening a KVM, the agent takes a screenshot, clicks the button, and carries on.

   your machine                                          machine being controlled
  ┌─────────────────────────────────┐                   ┌──────────────────────────┐
  │ Claude Code ── MCP ──► rdc mcp  │   Tailscale       │  rdc serve               │
  │ or:                    rdc -t … ├──── HTTP ────────►│  identifies you via      │
  │                                 │   (WireGuard)     │  tailscaled whois, then  │
  └─────────────────────────────────┘                   │  screenshots · input     │
                                                        │  windows · clipboard     │
                                                        └──────────────────────────┘

• No passwords, tokens or certificates. rdc serve listens only on the machine's Tailscale address and asks the local tailscaled who each caller is. You allow tailnet logins, device names or tags. Your tailnet is the security boundary.
• Scoped access and an audit trail. Each allowed identity can be limited to view, input or clipboard, and every request or rejection is written to a JSON-lines audit log.
• No shell. rdc is a screen-and-input surface only. Use SSH for commands.
• One binary, written in Rust. The same executable is the daemon, the CLI and the MCP server.

Status

CI builds and tests all three on every push. Tagged releases attach binaries, which are not code-signed (see Install).

Five-minute start

1. Install rdc on both machines. Download a release archive or cargo install --git https://github.com/bscott/rdc. Details and build dependencies: docs/install.md.

2. On the machine you want to control, write the allowlist into the config file, check readiness, and start the daemon:

# Linux: ~/.config/rdc/config.toml   macOS: ~/Library/Application Support/rdc/config.toml
# Windows: %APPDATA%\rdc\config.toml
[serve]
allow = [
  "you@example.com",                        # full control: tailnet login, device name, or tag
  { who = "monitor-bot", can = "view" },    # optional: limit an identity to screenshots
]

rdc doctor
rdc serve

If your Tailscale policy is not the permissive default, also allow the traffic there. For the config above, with the controlled machine tagged tag:rdc-host and the monitor device tagged tag:monitor:

"grants": [
  { "src": ["you@example.com", "tag:monitor"], "dst": ["tag:rdc-host"], "ip": ["tcp:7770"] },
]

The configuration guide maps every grant example to its policy rule, in both grants and legacy acls syntax. rdc doctor tells you if a permission or tailscaled is missing. When it works, install it as a background service so it survives reboots: rdc service install (macOS LaunchAgent or Linux systemd user service). The service reads the same config file, which is why the allowlist goes there rather than on the command line. macOS needs two one-time permission grants; follow docs/setup-macos.md.

3. On your laptop, talk to it by Tailscale hostname:

rdc -t studio-mac whoami                        # the daemon confirms who you are
rdc -t studio-mac shot --max 1568 -o shot.png   # screenshot, downscaled
rdc -t studio-mac click 640 400                 # desktop coordinates
rdc -t studio-mac key cmd+q

4. Give it to your agent. Add a target to your own machine's config file (same per-platform paths as above) and register the MCP server in Claude Code:

[targets.studio-mac]
url = "http://studio-mac.example-tailnet.ts.net:7770"

{ "mcpServers": { "studio-mac": { "command": "rdc", "args": ["mcp", "--target", "studio-mac"] } } }

The agent now has screenshot, click, type, key, focus and friends. It passes pixel coordinates from the screenshot it just saw; rdc converts them to desktop points. Every action returns a fresh screenshot so the agent can check its work. Full tool reference: docs/mcp.md. An Agent Skill is included that teaches agents how to use rdc well.

Documentation

The same pages are published as a site at bscott.github.io/rdc.

How it stays safe

rdc serve refuses to bind anything but a Tailscale address (or 127.0.0.1 with the explicit --dev-loopback testing flag, which disables auth). For each request it checks that the Host header names this machine, resolves the peer IP through tailscaled's whois, and matches the login, node name or tags against your grants. Tagged devices are identified by their tags only, never by the login of whoever created them. An empty allowlist refuses to start. Grants can be limited to view, input or clipboard, and everything is written to an audit log you can read with rdc audit. Read SECURITY.md before exposing a machine you care about.

License

rdc is free software under the GNU General Public License v3.0 or later. Use it, change it, ship it; if you distribute a modified version, publish your changes under the same terms so improvements stay open.