Sometimes you don't want full GUI automation — just a few safe, bounded actions on a machine. A lightweight control agent exposes those over a simple HTTP API.
Typical capabilities
| Capability | Example |
|---|---|
| Stats | CPU / RAM / GPU / disk telemetry |
| Media | Play / pause / next / volume |
| Power | Sleep / lock / shut down |
| Clipboard | Read / write remotely |
A small service listens on a chosen port and is surfaced through your assistant (e.g. as a tool), so it can answer "what's the GPU temp?" or "pause the music."
Fixed-capability vs full automation
- Control agent — a fixed menu of safe actions. Predictable, low-risk; use for everyday control.
- Computer use — arbitrary GUI tasks. More powerful, more caution.
Lock it down: keep the agent on the private mesh (Tailscale Network Topology), never the open internet, and limit it to non-destructive actions unless you really mean it.
Related: AI Computer Use and GUI Automation · Personal AI Assistant Architecture · Tailscale Network Topology