AI agents have access to files, terminals, and a handful of APIs with keys. But 99% of the internet requires a browser session.
brosearch gives agents direct access to the web through your already-logged-in Chrome. No API keys. No re-login. No scraping. The browser executes your session credentials, the same way you do it yourself.
agent calls: brosearch eval --js "return document.title"
→ runs JS in your Chrome tab → returns result as JSON
AI Agent (Claude Code / Cursor / OpenClaw / any CLI)
│
│ Python CLI (brosearch eval / fetch / navigate / detach / ...)
▼
┌─────────────────────────────────────────────────────┐
│ Daemon (Node.js, :19824) │
│ │
│ Adapter Router ──── loads adapters/*.js │
│ │ │
│ Extension Bridge ──── SSE ───────────────────────► │──► Chrome Extension
└─────────────────────────────────────────────────────┘ │
chrome.debugger (CDP)
│
┌───────▼────────┐
│ Your Chrome │
│ Real session │
│ Real cookies │
└────────────────┘
Key insight: The Chrome extension keeps a persistent SSE connection to the daemon. When the agent calls brosearch eval, the JS is sent to the extension over SSE, executed inside the real browser tab via chrome.debugger (CDP), and the result is returned as JSON. The website never sees a bot — it sees you.
| Playwright | Scraping | brosearch | |
|---|---|---|---|
| Browser instance | New headless browser | No browser | Your real Chrome |
| Login state | None — must re-login | Cookie extraction (fragile) | Already there |
| Anti-bot detection | navigator.webdriver=true |
IP blocks, CAPTCHAs | Invisible — it IS you |
| Complex auth (2FA, OAuth) | Very hard to replicate | Nearly impossible | Page handles it |
| Speed | Slow (browser launch) | Varies | Fast (tab already open) |
git clone https://github.com/haibindev/brosearch.git
cd brosearch
# One-line setup (installs Python package + builds Node daemon)
powershell -ExecutionPolicy Bypass -File scripts\setup.ps1
# Load Chrome extension manually:
# 1. Open chrome://extensions/
# 2. Enable "Developer mode" (top-right)
# 3. Click "Load unpacked" → select packages/extension/
# Start daemon in background
python -m brosearch daemon -b
# Verify
python -m brosearch doctorgit clone https://github.com/haibindev/brosearch.git
cd brosearch
# One-line setup
bash scripts/setup.sh
# Load Chrome extension (same manual steps as above)
# Start daemon in background
python -m brosearch daemon -b
# Verify
python -m brosearch doctorWSL only needs the Python CLI. The daemon and Chrome extension run on the Windows side.
# ── Windows side (PowerShell) ──
# Follow the "Windows" steps above to install daemon + extension
# ── WSL side ──
git clone https://github.com/haibindev/brosearch.git ~/brosearch
cd ~/brosearch
pip install -e . # Python CLI only, no Node needed
# Verify (auto-detects Windows host IP)
python -m brosearch doctorFirewall: If WSL can't reach the daemon, allow port 19824:
# Windows PowerShell (Admin) netsh advfirewall firewall add rule name="brosearch" dir=in action=allow protocol=TCP localport=19824
| Command | Description | Chrome |
|---|---|---|
eval --js <code> [--tab <pattern>] |
Execute JS in browser tab | ✅ |
navigate <url> [--tab <pattern>] |
Open URL in Chrome | ✅ |
fetch <platform/command> [key=val] |
Run platform adapter | ✅ |
search <query> [--engines ...] [--limit N] |
Multi-engine web search | ❌ |
read <url> |
Full-text via Jina Reader | ❌ |
capture [--tab <pattern>] [--duration N] |
Record network requests | ✅ |
generate --capture <file> --platform <p> --command <c> |
Generate adapter from capture | ❌ |
auto-generate --url <url> --platform <p> --command <c> |
Fully automatic: open→interact→capture→generate | ✅ |
console [--tab <pattern>] [--clear] |
Read console.log from tab |
✅ |
errors [--tab <pattern>] [--clear] |
Read JS exceptions from tab | ✅ |
detach [--tab <pattern>] [--all] |
Detach debugger (removes debug banner) | ✅ |
adapters |
List available adapters | ❌ |
daemon [-b] [--stop] [--status] |
Start/stop/check daemon | ❌ |
doctor |
Check daemon + extension health | ❌ |
# Execute JS in active tab
python -m brosearch eval --js "return document.title"
# Execute JS in a specific tab
python -m brosearch eval --tab "*://github.com/*" --js "return document.title"
# Navigate to a URL
python -m brosearch navigate "https://github.com"
# Run platform adapter
python -m brosearch fetch zhihu/hot
python -m brosearch fetch twitter/search query="Claude agent"
# Multi-engine web search (no Chrome needed)
python -m brosearch search "AI agent frameworks"
# Detach all debugger sessions (removes "debugging this browser" banner)
python -m brosearch detach --all- i18n: Automatically displays in English or Chinese based on Chrome language
- Badge: No badge when connected; red ✗ badge when daemon is disconnected
- Popup: Click the extension icon to see connection status
| Status | Meaning |
|---|---|
| 🟢 Ready | Daemon connected, extension active |
| 🟡 Reconnecting | Daemon running but extension needs refresh |
| 🔴 Daemon off | Daemon not running — start it first |
| Platform | Commands | Notes |
|---|---|---|
| Twitter / X | search, timeline |
Bearer + CSRF (Tier 2) |
hot |
Public JSON API; sub= param for subreddit |
|
| GitHub | trending |
Public page + DOM parsing |
| HackerNews | top |
Firebase public API |
| arXiv | search |
Public Atom API; no tab needed |
| StackOverflow | search |
StackExchange public API; no tab needed |
| npm | search |
registry.npmjs.org; no tab needed |
| Product Hunt | trending |
GraphQL (logged-in recommended) |
| V2EX | hot |
Public API |
| Platform | Commands | Notes |
|---|---|---|
| 知乎 Zhihu | hot, search |
Cookie (Tier 1) |
| 微博 Weibo | hot, hot-search |
Cookie; hot-search = 热搜榜 |
| 小红书 XHS | search |
Webpack injection (Tier 3) |
| B站 Bilibili | hot |
Cookie (Tier 1) |
| 36kr | hot, newsflash |
Cookie (Tier 1) |
| 掘金 Juejin | hot |
Token (Tier 2) |
| 雪球 Xueqiu | hot |
Cookie; market=CN/US/HK |
| 豆瓣 Douban | hot-movie |
Cookie; type=movie/tv |
Custom adapters go in
adapters/custom/(gitignored). Generate them withauto-generate.
One JS file per command:
// adapters/mysite/feed.js
module.exports = {
description: 'Get latest posts from MyForum',
tabQuery: { url: '*://myforum.com/*' }, // which tab to run in
buildJs: (args) => `
// Runs inside your Chrome tab. Cookies are already there.
const res = await fetch('/api/posts?limit=${args.limit || 20}', {
credentials: 'include'
})
const data = await res.json()
return data.posts.map(p => ({ id: p.id, title: p.title, url: p.url }))
`
}Three auth tiers:
| Tier | Method | Example |
|---|---|---|
| 1 | fetch() + credentials: 'include' |
Reddit, GitHub, V2EX |
| 2 | Extract Bearer/CSRF from cookies/globals | Twitter, Zhihu, Juejin |
| 3 | Access webpack module / global store | Xiaohongshu |
ANTHROPIC_API_KEY=sk-... python -m brosearch auto-generate \
--url "https://producthunt.com" \
--platform producthunt \
--command trendingThe AI interaction loop:
- Opens URL in Chrome
- Gets Accessibility Tree snapshot (
@ref role "name") - Decides actions:
capture,scroll,click @ref,type text [@ref],key-press Enter - Captures triggered API requests (with response bodies, up to 256KB each)
- Generates adapter JS from captured traffic
- Checks
console.log/ JS errors for self-diagnosis if generation fails
brosearch is inspired by and builds on the ideas from bb-browser — a brilliant project that pioneered the "your browser is the API" concept for AI agents.
brosearch is a lighter, self-contained version optimized for OpenClaw / Python-based agent setups, with a focus on Chinese platforms and fully automated adapter generation.
- Open an issue for bugs or new platform requests
- PRs welcome — especially new adapters
- Star the repo if it's useful
haibindev.github.io — personal site & blog
MIT © haibindev
brosearch 把已登录的 Chrome 浏览器变成 AI agent 可以直接调用的数据 API。
无需 API Key，无需重新登录，无需爬虫——Chrome 标签页本身就是认证凭据。
Python CLI → Node.js Daemon (:19824) → Chrome Extension (SSE+CDP) → 浏览器标签页
- Chrome 扩展通过 SSE 保持与 daemon 的长连接
- Agent 调用 CLI → daemon 转发命令 → 扩展在真实标签页中执行 JS → 返回 JSON
- 支持 i18n（中/英文自适应）
Windows:
git clone https://github.com/haibindev/brosearch.git
cd brosearch
powershell -ExecutionPolicy Bypass -File scripts\setup.ps1
# 加载 Chrome 扩展：chrome://extensions/ → 开发者模式 → 加载已解压 → packages/extension
python -m brosearch daemon -b # 后台启动 daemon
python -m brosearch doctorWSL（连接同机 Windows Chrome）:
# Windows 端：按上面步骤安装 daemon + 扩展
# WSL 端：只装 Python CLI
pip install -e ~/brosearch
python -m brosearch doctor # 自动检测 Windows 宿主 IP防火墙放行：
netsh advfirewall firewall add rule name="brosearch" dir=in action=allow protocol=TCP localport=19824
# 在浏览器中执行 JS
python -m brosearch eval --js "return document.title"
# 导航到指定页面
python -m brosearch navigate "https://zhihu.com"
# 运行平台适配器
python -m brosearch fetch zhihu/hot
# 多引擎搜索（无需 Chrome）
python -m brosearch search "AI agent"
# 移除调试横幅
python -m brosearch detach --all
# Daemon 管理
python -m brosearch daemon -b # 后台启动
python -m brosearch daemon --status # 查看状态
python -m brosearch daemon --stop # 停止内置 17 个平台，20+ 命令，包括：知乎/微博/小红书/B站/36kr/掘金/雪球/豆瓣（国内），Twitter/Reddit/GitHub/arXiv/StackOverflow/npm（国外）。
auto-generate 命令：AI 自动打开页面 → Accessibility Tree 分析 → 交互（点击/滚动/输入）→ 抓取 API → 生成适配器 JS。
灵感来源和致谢：bb-browser。
个人主页：haibindev.github.io 微信公众号：海滨code
