Automating a Godot game with Claude Code via Parsec
How to give Claude Code the ability to play a game running on a remote machine over Parsec, using eyehands' pointer-lock-compatible input.
Guides, deep dives, and release notes — mostly about eyehands and Claude Code running real workflows on Windows.
Task-oriented walkthroughs for specific jobs.
How to give Claude Code the ability to play a game running on a remote machine over Parsec, using eyehands' pointer-lock-compatible input.
A practical guide to teaching Claude Code to operate native Windows applications using UI Automation, OCR, and the eyehands local HTTP server.
Legacy Windows apps without Selenium bindings, test frameworks, or official APIs can still be tested. Here's how to drive a WinForms tool with Claude Code via eyehands.
Windows UI Automation is the accessibility API screen readers use — and the cleanest way to find controls programmatically. Here's how to use it from Python without getting lost in COM.
Five real recipes for letting Claude Code drive your Windows desktop via eyehands: inbox triage, log scraping, dashboard screenshots, deploy health checks, and form auto-fill.
Why eyehands uses SendInput under the hood, which games it actually works with, and a walkthrough of building a simple automation bot with safe input throttling.
Turn eyehands into a lightweight WinForms UI-test harness: click elements by name via UI Automation, assert on screen content via OCR, and keep tests deterministic with click_and_wait.
Technical explainers. What, why, and how much it actually costs.
Why an HTTP server is the right shape for AI agent tools, when it isn't, and design decisions I made building eyehands.
Measured token costs for the same Windows automation task done screenshots-first vs UIA-first. The 4× difference, and where it comes from.
If Claude is waiting for a build to finish, it shouldn't be paying for screenshots while nothing changes. Frame-hash polling with If-None-Match is how you get wait-for-change for free.
Three ways an AI agent can see a Windows desktop, with real token costs, failure modes, and when to use each.
Why most Python Windows automation scripts get wrong coordinates on high-DPI monitors, and the single DllCall that fixes it.
Most Python automation libraries use keybd_event and mouse_event under the hood. These APIs silently fail inside pointer-lock applications. Here's why, and the SendInput fix.
Release notes and the eyehands privacy model.
A focused pass on security: DNS rebinding defense, constant-time token comparison, request size caps, HttpOnly session cookies, and more.
1.6.0 adds an anonymous heartbeat so free users show up in stats. Full disclosure of what's sent, what isn't, and how to opt out.
Full disclosure of eyehands 1.6.0's anonymous install heartbeat: exact payload, what's excluded, and three ways to opt out.