huashu-chrome MCP
An MCP server and Chrome extension give any agent 22 tools for your own logged-in browser, verifying the effect of every action.
- 593 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent works under your session on any site: it reads pages, calls their APIs with your cookies and types data.
- Form filling and uploads happen in working portals where a mistake writes data to customers or a CRM.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/huashu-chromeDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent works under your session on any site: it reads pages, calls their APIs with your cookies and types data.
- Form filling and uploads happen in working portals where a mistake writes data to customers or a CRM.
This is third-party code. Review the repository files before installing.
What it does
huashu-chrome ties an agent to your everyday Chrome. The npm package holds a CLI with `install`, `mcp` and `doctor` commands, an MCP server as the agent's interface, a local bridge on 127.0.0.1 and a Chrome extension that performs the actions. The agent works under your site logins, so it needs no API keys or re-authentication. Twenty-two tools are arranged in three layers of information. Data layer: `network` shows a page's API calls, `fetch` calls them with your cookies and pages through everything into JSONL, `download` pulls files. Action layer: `snapshot` photographs interactive elements with refs, `fill` fills a form in one call, plus `click`, `type`, `select`, `key`, `upload`, `read_text` and `query`, while `act` runs a chain of steps and verifies each. Human layer: `ask` hands you a captcha, a QR login, an SMS code and confirmations. Every write operation reports what changed on the page and warns outright when the page did not react. Payment, submit and delete are never done automatically inside an `act` chain. Built-in notes for twenty-three sites (12306, JD, Bilibili, Feishu, X, Zhihu and more) record working endpoints and dead ends, and the `learnings` tool adds your own.
Who it is for. Users of Claude Code, Codex, Cursor and other MCP clients who want an agent inside back offices and CRMs where they are already logged in, without a separate browser.
Good fit when
- Thirty rows from a CSV have to go into a CRM open in your browser.
- A paginated list from a site must be exported to JSONL using your session.
- An agent needs a multi-step form filled, with the result checked after every step.
- A partner portal asks for a captcha or an SMS code and a person must do it while the agent waits.
- Several agents work at once and need separate tabs without grabbing each other's pages.
Not a fit when
- You need a clean browser without your login for public scraping: the server gives an agent access to your accounts.
- The page is protected: the Chrome Web Store and similar internal pages are out of reach of the extension.
- The agent must not see the content of your accounts: it receives page snapshots and text.
Example request
Take the CSV with thirty customers and enter them into the CRM open in my Chrome, confirming the result after each entry.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
Node.js 20 and a Chrome extension are required; the extension is installed by hand because the browser does not let a script do it. The `install` command detects 20 known agents and writes their configuration, while OpenCode must be configured manually.
Actions are harder inside closed shadow DOM and canvas, and the extension cannot work on browser-protected pages. On x.com React does not render in background tabs, so actions fail there.
Rules update with the package, and the site notes hold as of their check date and may go stale.
How to disable. Remove the huashu-chrome entry from your client's MCP configuration (for Claude Code `claude mcp remove huashu-chrome`) and disable the extension in `chrome://extensions`. The command log lives in `~/.huashu-chrome/audit.jsonl`.
FAQ
How do I connect huashu-chrome to Claude Code?
`npx huashu-chrome install` finds agents and writes the configuration itself. To do it by hand run `claude mcp add huashu-chrome -- npx -y huashu-chrome mcp --client claude-code`. You must load the extension yourself in `chrome://extensions`.
How do I check that it works?
Run `npx huashu-chrome doctor`. The message to look for says the handshake is fine and the Chrome extension is online. The bridge starts by itself on the agent's first call.
What about captchas and payments?
The `ask` tool hands captchas, QR logins and SMS codes back to you. A payment confirmation appears as a card in the browser: it lives in the extension, out of the agent's reach. With no answer the action is declined.
Where are the agent's actions recorded?
In the log `~/.huashu-chrome/audit.jsonl`. Typed text is masked and passwords are detected by field type. Read the log with `npx huashu-chrome audit`.
What about OpenCode?
Automatic setup is not supported for it yet. Add an `mcp` entry to `opencode.json` with the command `npx -y huashu-chrome mcp --client opencode`.
Discussion
No questions or reviews yet.
Related
MCP serversA CLI and MCP server that lets an agent drive your real Chrome with its logged-in sessions: 103 commands across 36 platforms plus browser automation
A remote MCP server and Chrome extension: the agent reads data behind your logins and operates the sites you already use
An agent drives your real logged-in Chrome: a CLI, MCP and an extension with a separate tab group per session
A Chrome extension with an MCP server: the agent drives your everyday browser with its sessions, tabs, bookmarks and history
An MCP server that drives your already open Chrome and stops to ask you about a 2FA code, a CAPTCHA, or a choice only you can make
Chrome control for AI agents: a local Go HTTP server with a CLI, MCP, profiles and a skill for Claude Code
README in short
The Chinese README states the principle that an action counts only if the page changed, shows sample tasks and a comparison table with six alternatives, then install, tool groups, effect verification, the batch `act` call, accumulated experience and nine security measures. Privacy and troubleshooting close it.