Eyehands

eyehands

AI agents on Windows are slow because every screenshot round-trips through the cloud. Claude Computer Use takes 2-4 seconds just to see the screen. eyehands runs locally and does it in 57ms.

Download the Windows .zip from [fireal.dev/eyehands/](https://fireal.dev/eyehands/) or [GitHub Releases](https://github.com/shameindemgg/eyehands/releases), extract, and run:

eyehands.exe

That's the whole setup. HTTP server on `localhost:7331`, works from any language:

curl http://localhost:7331/screenshot --output screen.jpg
curl -X POST http://localhost:7331/click_at -d "{\"x\":500,\"y\":300}"

Or use the Python SDK (`pip install eyehands-sdk`):

from eyehands import EyehandsClient
c = EyehandsClient()
c.screenshot("screen.jpg")       # 57ms
c.click_at(500, 300)             # instant
c.type_text("hello world")      # native SendInput, not clipboard paste

	eyehands	Claude Computer Use
Screenshot	57ms	2-4s
5-action batch	67ms	3-5s
Resolution	1920x1080	1456x816
OCR text search	Optional*	No
UI Automation	Yes	No
Games / pointer-lock	Yes	No

\* OCR (`/find`) requires ~2GB of EasyOCR/PyTorch dependencies, not bundled in the standalone .exe. Install from source or add easyocr to system Python to enable. All other endpoints work out of the box.

7-day free trial, then $49 one-time at [fireal.dev](https://fireal.dev). No subscription.

How it works

A background thread captures the screen at ~20fps into a rolling buffer. The capture backend picks the fastest option at startup: BetterCam (DXGI, ~120fps) if available, then DXcam (~39fps), then mss (GDI fallback). Mouse and keyboard go through `SendInput` via ctypes, which hits the Raw Input pipeline. This matters because most Python automation tools use the deprecated `keybd_event`/`mouse_event` APIs, and pointer-lock apps (games, Parsec, full-screen browsers) silently ignore those.

There's also OCR via EasyOCR (resu