olmatter/glm-coding-helper

This project is a browser automation tool for people trying to buy a limited-availability AI coding subscription called GLM Coding Plan, sold by the Chinese AI company Zhipu. Because the subscription sells out in seconds each morning, the author built a Tampermonkey userscript that works inside Chrome or Edge to automate the checkout steps and handle the obstacles that slow most users down. The core problem the tool addresses is a Chinese character-click captcha that appears during checkout. Instead of sending captcha images to an online service, the script sends them to a local backend running on your own machine. That backend uses two open-source recognition libraries, PaddleOCR and YOLO, to identify which characters to click and then tells the browser where to click. This keeps your captcha images off third-party servers entirely. Setup involves two parts. First, you install the Tampermonkey browser extension and add the userscript from Greasy Fork or the downloaded file. Second, you run a local backend by double-clicking a startup script included in the release package. The backend listens on your machine at a local address, no internet connection is needed for the OCR step once the backend is running. The release comes in two flavors: a portable package that includes the full Python environment, and a smaller online installer that downloads the environment on first run. The script supports running multiple browser windows at the same time, which increases the chance of getting through during the high-traffic window around the sale start time. Configuration options include whether to automatically confirm the captcha, whether to auto-close invalid payment dialogs, and which subscription tier to try first. By default the script takes a conservative approach, clicking captcha characters but leaving final confirmation to the user. The project is built on top of an earlier Greasy Fork script by another author and is released under the GNU GPLv3 license. The accuracy data in the README shows the local OCR pipeline reaching 100 percent on both a hidden test set and a larger stress test of 379 captcha images, compared to around 96 percent for the lighter alternative approach.