Allow your claws to do things remotely on a Desktop machine via MCP
Full control of a remote machine via Remote Claws MCP: screenshots, mouse/keyboard, browser automation, run commands, read/write files on the remote host.
PAUL GRESHAM
@wentbackward
Install
$ openclaw skills install @wentbackward/remote-clawsRemote Claws — Remote Machine Control
Provide permission-based access to a remote desktop machine via the remote-claws MCP server (https://github.com/wentbackward/remote-claws) allowing your agent to run in a sandbox.
Four tools, each taking an action parameter: remote_browser, remote_desktop,
remote_exec, remote_files. Read each tool's description for its action
list — an unknown action returns the valid list.
CRITICAL: Remote vs Local
The remote_* tools act on the REMOTE machine. OpenClaw's built-in browser,
exec, read/write/edit act on the LOCAL gateway machine. Never
substitute one for the other. Take care not to confuse remote-claws with remote
SSH commands.
Strategy
- Screenshot first.
remote_desktop(action="screenshot")before clicking or typing; use the returned coordinates to target actions. Re-screenshot after actions — windows move, dialogs appear. - Prefer remote_browser for web tasks. CSS selectors are
resolution-independent. Only fall back to
remote_desktopfor things the browser can't reach (native dialogs, file pickers). - Prefer element names over coordinates.
remote_desktop(action= "click_element", ...)targets controls by name — survives window moves. - Exec is async.
remote_exec(action="run", ...)returns a process_id; poll withaction="get_output"(wait=true blocks),action="send_input"for stdin,action="kill"when done. - Denied actions are final. A "permission denied" result means server policy — do not retry unless the user requests it.
Common actions
- Desktop: screenshot, mouse_click, mouse_move, mouse_drag, scroll, type_text (ASCII only), press_key, find_window, focus_window, list_elements, click_element, get_element_text
- Browser: navigate, click, fill (clears first, Unicode-safe), type (appends), press_key, get_text, get_html, eval_js, screenshot, wait_for, select_option, go_back, go_forward, tabs_list, tab_new, tab_close
- Exec: run, get_output, send_input, kill, list
- Files: read (base64, offset/limit for chunks), write (base64), list, delete, move, info
Authentication & Security
The remote-claws MCP server requires a bearer token, configured in
openclaw.json when registering the server. Unauthenticated connections get
401. The server also supports IP allowlisting (allowed_ips), host header
validation (allowed_hosts), and per-action permission policies
(permissions.json). See the setup guide
and README.
Important Notes
- Screenshots are JPEG, max 1280x960. Coordinates are absolute pixels. If your main
LLM supports images, openclaw media handling should just work. Otherwise use
remote_desktop(action="screenshot", save_to_disk=true)and download the binary. - Reading Remote Files:
remote_files(action="read", ...)returns base64 only up to a threshold. Bigger reads fail with"inline read would return N bytes"to avoid context token explosion. Re-issue the same read withas_url=trueand use the returned URL to download — do NOT retry the plain read. Small chunks via offset/limit will work. type_textis ASCII only. For Unicode, use browserfill, or clipboard:remote_exec(action="run", command="powershell", args=["Set-Clipboard", ...])thenremote_desktop(action="press_key", keys="ctrl+v").- The browser launches on first use and stays open across calls. Sessions persist (cookies, local storage).
Top skills in this category
API Gateway
@byungkyuCall third-party APIs through the Maton gateway, which injects the credential for an app the user has already connected. Use this skill when the user names a connected app and a concrete action in it - read a mailbox, query a CRM, file an issue, update a spreadsheet, run a query through a connected
Playwright MCP
@spiceman161Browser automation via Playwright MCP server. Navigate websites, click elements, fill forms, extract data, take screenshots, and perform full browser automation workflows.
Computer Use
@ram-raghav-sFull desktop computer use for headless Linux servers. Xvfb + XFCE virtual desktop with xdotool automation. 17 actions (click, type, scroll, screenshot, drag,...
Local Places
@steipeteSearch for places (restaurants, cafes, etc.) via Google Places API proxy on localhost.
OneDrive Files
@hith3shBrowse, search, download, and share OneDrive files, create folders, upload files, and manage file actions via Microsoft Graph. Use this skill when users want...