anzal1

Fasthands — AI skill for Claude Code

AI community

Open-source, model-agnostic computer use: diff observations, guarded action batching, and xray — the HTML contract every agent ignores.

How to install Fasthands

This entry records only its repository, not the path inside it, so there is no exact command to give. Open anzal1/fasthands and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Fasthands does

Open-source, model-agnostic computer use: diff observations, guarded action batching, and xray — the HTML contract every agent ignores. 91% fewer tokens than a screenshot loop, benchmarked and reproducible keyless.

Alternatives in AI

  • Vllm Mlx — High-performance OpenAI and Anthropic compatible LLM inference server for Apple Silicon 1.5k ★
  • Claude Skill Aso Appstore Screenshots — A Claude Code skill that generates high-converting App Store screenshots for your iOS app 1.2k ★
  • Ppt Agent Skill — 🎨 World-class AI presentation generator · 26 styles · 18 charts · benchmarked against Linear/Anthropic/Stripe 139 ★

README

fasthands

**Open-source, model-agnostic computer use. Same task, fewer tokens, fewer turns, with any model as the brain, including a 4B running on your laptop.**

Computer-use agents today come in two flavors. Anthropic's reference loop sends a screenshot every turn (about 1,366 image tokens at 1280x800) and takes one action per turn. OpenAI's GPT-6 Astra loop is smarter: accessibility trees and code actions, but it resends a fresh full tree every turn, it is closed, and it runs one company's models. Its own write-up admits stale trees and state drift between observing and acting as open failure modes.

fasthands attacks the loop, not the model. Every claim below is measured, reproducible with zero API keys, and reported with its caveats.

The claims, and where each one is proven

# Claim Evidence Reproduce
1 91.3% fewer observation tokens than a screenshot loop, 61.2% fewer than an Astra-style full-tree loop Proof 1, 120 runs npm run bench -- --repeat 5
2 Blind batching misclicks catastrophically under DOM drift; guarded batching never does Proof 2, 40 trials: 0% vs 100% node --experimental-strip-types src/bench/drift-test.ts
3 Diff observations stay flat while every other strategy grows with page size Proof 3, 38 tokens vs 92,732 at N=2000 node --experimental-strip-types src/bench/scale-test.ts
4 Reading the HTML contract (xray) removes whole turns of error recovery Proof 4, signup trap: 2 turns vs 3 to 8 included in the bench
5 The harness carries a real 4B local model through real tasks Proof 5, 46% fewer live tokens, trap flips fail to pass --provider compat --review
6 Canvas, WebGL and drawing are first-class: observe below the DOM, act in stro