Settled Computer — AI skill for Claude Code

settled-computer

**Fast computer use for AI agents.** If your computer-use agent types `time.sleep(2)` between every click, re-screenshots after every action, and burns vision tokens re-reading an unchanged screen — that's the problem this fixes.

An MCP server for desktop computer use with **event-driven settling**: every action waits until the screen actually reacts and stops changing, then returns the settled screenshot and a one-line verdict. No fixed sleeps, no separate screenshot round trips, no model-side guessing about *when* to look.

Built for agent loops (Hermes, Claude Desktop, Claude Code, any MCP host). The core idea: **the server answers "did the UI react, and has it stopped?" at 20 Hz with a ~1 ms frame diff, so the model doesn't need an extra vision turn just to find out when it is safe to look.** Whether the reaction was the *right* one is still the model's job: judge by the image.

Why your computer-use agent is slow (and how this fixes it)

Naive computer-use loops look like this:

click -> guess a sleep -> screenshot (full vision turn) -> hope

Timings are guesses: too short means acting on half-loaded UIs, too long means wasted seconds. Each "is it ready yet?" costs a model turn and a screenshot that enters context.

settled-computer replaces that with:

click -> server watches the screen -> returns the AFTER frame + verdict

• Reacted and settled in 0.4 s → Screen settled 0.40s after the action.
• Nothing changed → No visible change… (it may simply have had no visible effect: an already-focused field, an already-set state). If the model received an image within the last 45 s and the screen is unchanged, the image is omitted to save vision tokens.
• One region keeps moving (video/spinner) → bails at ~1.5 s with the region's bounding box instead of stalling 8 s, and after two consecutive bails the region is auto-ignored.

Tools

tool	what it does
act	The main one. Runs 1–8 actions in O