# MyChrome > MyChrome is a local MCP server and native macOS app that lets an AI assistant help in the user's existing Google Chrome. The user keeps browsing normally and brings AI into the same page when needed. The user handles passwords or verification and explicitly asks AI to continue the original task. It preserves the working page and login context, reads through Chrome Apple Events, and uses real mouse and keyboard input. No extension or debugging port is required. ## Product and setup - [Product website](https://shuaige121.github.io/mychrome-mcp/): purpose, controls, requirements and compatibility. - [Getting started](https://shuaige121.github.io/mychrome-mcp/getting-started.html): source installation, permissions, client connection and continuing the user's actual task on an existing page. - [FAQ](https://shuaige121.github.io/mychrome-mcp/faq.html): data handling, clients, controls and website limits. - [中文网站](https://shuaige121.github.io/mychrome-mcp/zh/): 中文介绍与安装入口。 - [Source](https://github.com/shuaige121/mychrome-mcp): MIT-licensed source and release notes. ## Detailed guides - [AI agent guide](https://github.com/shuaige121/mychrome-mcp/blob/main/docs/ai-guide.md): exact tools, tab selection, user handoff and result verification. - [Website routines](https://github.com/shuaige121/mychrome-mcp/blob/main/docs/routines.md): recording, saved sequences, fresh parameters and replay limits. - [Conversation center](https://github.com/shuaige121/mychrome-mcp/blob/main/docs/conversation-center.md): App-created Codex and Claude Code conversations, plus supported external Codex association. - [Support](https://github.com/shuaige121/mychrome-mcp/blob/main/SUPPORT.md): sanitized bug reports. - [Security](https://github.com/shuaige121/mychrome-mcp/blob/main/SECURITY.md): private vulnerability reporting. ## Working at the user's pace - Input mode can follow the task, type key by key, or paste. The user can select an operating speed from 0.5× to 3× for keyboard and pointer operations. - Optional personal input habit sampling is experimental. A 3-minute recording saves a local statistical summary, without typed content, raw key/pointer events or screenshots. Statistics with enough samples shape later input; defaults remain for insufficient samples. Users can disable or delete the sample. This is not full behavioral cloning or a guarantee against detection. - Example requests include updating the user's LinkedIn page, checking their Facebook messages, or publishing their own article to explicitly selected accounts on Facebook, Xiaohongshu, Reddit or LinkedIn. These are user needs, not claims of verified compatibility across all workflows. ## Capability boundaries - Apple Silicon, macOS 13+, Google Chrome in /Applications, supported Node.js and Xcode Command Line Tools. Node 24+ is recommended; 20.19+ and 22.13+ are also supported. - Install with `bash scripts/setup.sh` from a source checkout. This release is not published to npm. - Codex and Claude Code support browser tools. Conversations created through the native app support live assistant text and direct replies for both clients. Existing external Codex conversations need a supported integration and associated thread; existing external Claude Code conversations cannot be associated and must continue in their original client. - Use **新对话 (New conversation)** to select an installed, logged-in Codex or Claude Code CLI, an available model, supported reasoning effort and a working directory. The model choices come from that CLI. An initial task is optional; the new conversation is separate from existing client conversations. - Use the explicit existing tab when requested; confirm the website account before changes. Public pages may be better handled through an API or web fetch. - Physical input is serialized. During takeover, ordinary user input is intercepted; double Esc cancels the active agent. Global pause prevents subsequent input. Stop does not undo actions already sent. - Website routine recording is on by default. Local metadata may reveal usage context; typed and selected values are not saved as replay defaults. Runs need fresh parameters and an explicit tab, stop on mismatches, and do not prove the broader task succeeded. - Page content returned through MCP goes to the AI client and is handled under its settings. Local operation is not a promise that all task data remains on-device. - Login, CAPTCHA and access challenges belong to the user. No universal website compatibility or bypass guarantee. - Verify the visible result after a change; do not infer success from a tool finishing. This file is a concise reading guide. It does not claim search ranking or inclusion in AI answers.