What is MyChrome?
MyChrome is a local MCP server and native macOS app. Keep using Chrome as usual and bring in AI on the same page when you need help. Take over when a step needs you, then tell it to continue. It uses your real pointer and keyboard, with adjustable pace and optional personal rhythm sampling.
Which computers and clients work?
Apple Silicon Macs running macOS 13+ with Google Chrome, supported Node.js and Xcode Command Line Tools. Codex and Claude Code support the browser tools; other local stdio MCP clients can use the generated connection paths. Codex and Claude Code conversations created in MyChrome support live text and direct replies in the native window. Existing external Codex conversations require a supported integration and associated thread. Existing external Claude Code conversations cannot be associated; continue them in Claude Code.
Click 新对话 (New conversation) to choose an installed, logged-in Codex or Claude Code CLI, an available model, supported reasoning effort and a working directory. Model choices come from the selected CLI. The first task is optional, and this creates a separate conversation.
Can it use my existing login?
Yes, it operates your installed Chrome. Ask your assistant to use the exact existing tab and confirm the account displayed by the website before making changes. A Chrome profile and a website account are not necessarily the same identity. Claude Code setup.
Does it work on every website?
What works depends on the site’s controls, account permissions and current page. You may need to take over for login or verification. The task examples on this website describe requests you can make; they do not establish compatibility with every workflow.
Can it use my input habits?
Choose an input mode and pace from 0.5× to 3×. The optional My input habits feature samples your rhythm for 3 minutes, then uses typing and pointer statistics with enough samples to shape later input. Sampling is experimental; you can turn it off or delete its data. Only a local statistical summary is saved, without typed content, raw paths or screenshots.
How does AI continue after I take over?
The handoff instructions tell the AI to wait for your request to continue. After handling the step, say “continue” in the original conversation; it should read the page again before proceeding. Page-change or input detection only ends the observation; your AI client must follow the handoff instructions.
How do I stop it?
Press Esc twice quickly, within 0.8 seconds, to cancel the current agent and return keyboard and mouse control. The agent must wait for a new instruction before taking over again. To prevent new input across agents, use Pause AI actions after control is returned. Stopping does not undo actions already sent.
Can I use the mouse while it works?
During takeover, ordinary physical input is intercepted. There is one pointer and keyboard on the desktop, and agent input takes turns. Reading can happen concurrently. Use double Esc when you need to take over.
What stays on my Mac?
MyChrome's settings, operation records and website routine metadata are local. Routine recording is enabled by default and can be turned off. It records filtered page/target descriptions and action order; typed and selected values are not saved as replay defaults. Paths, labels and routine names may still reveal usage context. Page content returned through MCP goes to your AI client and is handled under its settings. Do not share runtime directories or account screenshots.
What are website routines?
Saved continuous sequences of supported website actions. Review the steps, enter fresh values and select an existing tab for each run. Replay uses the same guarded input path without model calls between steps and stops when the page or target no longer matches. It does not support every control or prove an external business result. Full routine guide.
Is it free?
The project is open source under the MIT License. Your AI client supplies model access, and its own charges or subscription may apply. This release uses source installation and is not published to npm.
Should I use it for every page?
Choose the tool that fits the task. Public information often needs only an API or web fetch; headless browser automation suits JavaScript pages without your personal session. MyChrome is useful when your existing Chrome context or shared page matters. Choosing a browser tool.
Does it always save tokens?
There is no general savings guarantee. Page reads can return an outline or focused text instead of an image; actual model usage depends on the page, client and task. How page reading works.