My project: https://github.com/raghibrm/remote-pc-mcp
If you run agents (Claude Code, Cursor, Cline, anything MCP-compatible) and have more than one machine, this turns every box into something the agent can operate: run commands, manage files, watch CPU/RAM/GPU, launch and babysit long jobs, take screenshots, click and type when something needs a GUI touch.
Why it might fit this sub:
- Fully self-hosted. A small pip-installable Python package plus a supervisor, no cloud, no account, no telemetry, nothing phones home.
- Tailnet-first. Set
REMOTE_PC_MCP_HOST to your Tailscale IP and the listener is unreachable from anywhere outside your tailnet. The docs explicitly tell you not to expose it publicly without TLS and a reverse proxy.- Boring, auditable install. Windows: a Startup-folder shortcut. Linux: a systemd user unit (
install.sh --linger to survive reboots). --uninstall removes everything. The supervisor restarts the server with exponential backoff on crash.- Honest threat model. The bearer token is root-equivalent because
shell_exec runs arbitrary commands as the daemon user. That's the point of the tool, and the README says so in bold instead of burying it. Constant-time token comparison, sanitized error output.I run it on a GPU box and an old laptop that serves media, and dispatch heavy or long-running work to them from the machine I actually sit at.
v0.5.0, on PyPI and the official MCP registry, test suite + CI + CodeQL. Feedback welcome, especially on the security posture.
https://redd.it/1x0ehzs
@r_SelfHosted
