Disclosure: this post is from the RockDesk team. RockDesk is the tool the setup below uses.
Claude Code, Codex and Gemini CLI work well on the machine they run on. This guide lets them operate your other computers and devices the same way:
It works by machine ID, with nothing to open on the router, and takes about five minutes to set up. The other machine can be a Windows PC, a Mac, a Linux server without a screen, a robot or dev board running Linux, or an Android phone (screen, taps and typing).
The other machine can sit behind a home router, on the office network or on a mobile connection. The agent runs commands there the way it runs them locally and gets the real exit code back, so &&, || and if behave as they do over ssh. A local script can be piped in, and long jobs keep running after the call returns.
rds 123456789 shell -c "uptime; df -h /"
cat deploy.sh | rds 123456789 shell -c "bash -s"
rds 123456789 spawn "cd ~/proj && make -j8" # detached: prints the pid and its log file
On a Windows machine shell -c runs PowerShell, so the agent writes PowerShell there.
rds 123456789 mount :/home/me/project ~/remote-project
The remote project now looks like an ordinary local folder. The agent's usual file tools, searches, editors and builds work on it directly, and every change is written straight to the remote machine. When it is done: rds 123456789 umount ~/remote-project.
Linux mounts it with FUSE, macOS needs nothing extra, and on Windows the installer brings what is needed and the mount can be a drive letter (Z:). For one-off transfers there are rds <id> cp and rds <id> sync.
For apps that have no command line, the agent takes a screenshot, reads it, clicks and types, then takes another screenshot to check the result.
rds 123456789 screenshot -o shot.png
rds 123456789 click 812 440
rds 123456789 type "quarterly report"
rds 123456789 key Return
rds 123456789 forward 15432:db.internal:5432
A database, a web UI or a device that only the remote machine can reach is now on 127.0.0.1:15432 here. The host name is resolved on the remote side, so names that only exist on that network work. rds 123456789 unforward 15432 stops it.
There is more: clipboard, file copy and sync, serial ports, USB devices and a VPN to the remote network. Here it is on a Windows PC and on a Mac:
Every machine that runs RockDesk gets a 9-digit ID. The rds command line hands each command to the RockDesk service on your own machine, which keeps one encrypted connection per peer: direct when it can get through the NAT, through a relay otherwise. Nothing has to be opened on the router and the other machine needs no sshd, and that machine decides who may connect.
Download it from rockdesk.io. There are builds for Windows, macOS (Apple silicon and Intel), Linux (.deb, x86_64 and arm64) and Android. Install it on the computer your agent runs on and on every computer you want it to work on. New devices start with a free trial.
Note the 9-digit ID shown on the machine you want to control.
On a Linux machine without a screen (a server, a robot, a dev board):
sudo systemctl enable --now rockd # starts at boot
sudo rds status # prints this machine's ID
sudo rds otp # prints a one-time password
The skill is a single file that teaches the agent the rds commands. It also tells the agent to check that RockDesk is installed and to send you to the download page if it isn't.
Claude Code:
/plugin marketplace add rockdesk-io/rockdesk-skill
/plugin install rockdesk@rockdesk
Codex:
codex plugin marketplace add rockdesk-io/rockdesk-skill
codex plugin add rockdesk@rockdesk
Gemini CLI:
gemini extensions install https://github.com/rockdesk-io/rockdesk-skill
Any other agent that reads agent skills:
npx skills add rockdesk-io/rockdesk-skill
If you prefer MCP: rds mcp is an MCP server over stdio with 18 tools (shell, spawn, copy, ls, screenshot, click, type, key, scroll, clipboard, port forwarding and more). Screenshots come back as images.
claude mcp add -s user rockdesk -- rds mcp
codex mcp add rockdesk -- rds mcp
rds is not on the PATH on macOS and Windows. Use /Applications/rockd.app/Contents/MacOS/rds or C:\Program Files\rockd\rds.exe there.
The machine being controlled decides who gets in, and an ID on its own doesn't get anyone in. Pick whichever suits you:
sudo rds otp on a headless Linux box). It changes after each use.rds <id> auth once and it is remembered.rds key gen on your machine, then rds authorize "<the public line>" on the other one. After that you get in with no password.
Mounting folders, port forwarding and the other features can be switched on or off on the other machine.
Some things to try:
The first command to a machine takes a moment to connect. Later commands reuse the same connection and are almost instant.
Use whichever you like. The skill works in any agent that can run shell commands, and it also covers things that aren't MCP tools yet: mounting folders, file sync, the VPN, serial ports and USB. MCP gives you native tools in clients that support it, and screenshots arrive as images. You can install both.
Do both machines need to be on the same network? No. That's the point of the ID: it works across NATs and between networks, and falls back to a relay when a direct path isn't possible.
Does the other machine need a screen? No. On headless Linux use the commands from step 1.
Is there a web version? Yes, at rockdesk.io/web, for people. Agents need the rds command line, which comes with the desktop app.
What does it cost? New devices start with a free trial. See pricing.
If you try it, tell us what you'd like your agent to do on your other machines. We read every comment.