halley

fasthands

Computer use that sends diffs, not screenshots: 91% fewer tokens, any model as the brain

Area
Agents
Language
TypeScript
License
MIT
Plate
No. 03 of 14
Fig. 03fasthands
Move your pointer over it

What came back

  1. 91.3%

    fewer observation tokens than a screenshot-every-turn loop, over 120 verified runs.source

  2. 0 of 20

    wrong-button clicks when the page shifts mid-action, against 20 of 20 without the guard.source

  3. 39 tokens

    to observe a change on a 2,000-element page, instead of 92,732 to resend the whole tree.source

fasthands is an open-source computer-use engine that observes a page once and then sends only what changed. Every click is hash-checked at the moment it fires, so a page that shifts under the agent aborts the batch instead of hitting the wrong button. It runs with Claude, GPT, or a small model on your own laptop.

How it works

  1. 01

    The engine takes one compact tree of the interactive page, gives each element a stable ref and content hash, and afterwards sends only diffs.

  2. 02

    The model returns a JSON batch of actions; the executor re-hashes every ref before firing and aborts with a drift note if the page changed.

  3. 03

    xray adds the page's own validation rules to the observation, so the agent fixes forms before submitting instead of after a failed submit.

What it does

  • Observes the page as a compact tree once, then sends diffs with stable element refs
  • Guarded action batches: every ref is re-hashed before it fires and the batch aborts on drift
  • xray reads the HTML contract (required, pattern, min/max) and blocks a submit the browser says will fail
  • Reads canvas draw commands and WebGL scene graphs, and draws with guarded pointer strokes
  • Works with Anthropic, OpenAI or any OpenAI-compatible endpoint such as Ollama, with no SDK dependencies