You're already logged into Twitter, Reddit, YouTube, Zhihu, Bilibili, LinkedIn, GitHub — bun-browser lets AI agents use that directly.
bun-browser site twitter/search "AI agent" # search tweets
bun-browser site zhihu/hot # trending on Zhihu
bun-browser site arxiv/search "transformer" # search papers
bun-browser site eastmoney/stock "茅台" # real-time stock quote
bun-browser site boss/search "AI engineer" # search jobs
bun-browser site wikipedia/summary "Python" # Wikipedia summary
bun-browser site youtube/transcript VIDEO_ID # full transcript
bun-browser site stackoverflow/search "async" # search SO questions103 commands across 36 platforms. All using your real browser's login state. Full list →
The internet was built for browsers. AI agents have been trying to access it through APIs — but 99% of websites don't offer one.
bun-browser flips this: instead of forcing websites to provide machine interfaces, let machines use the human interface directly. The adapter runs eval inside your browser tab, calls fetch() with your cookies, or invokes the page's own webpack modules. The website thinks it's you. Because it is you.
| Playwright / Selenium | Scraping libs | bun-browser | |
|---|---|---|---|
| Browser | Headless, isolated | No browser | Your real Chrome |
| Login state | None, must re-login | Cookie extraction | Already there |
| Anti-bot | Detected easily | Cat-and-mouse | Invisible — it IS the user |
| Complex auth | Can't replicate | Reverse engineer | Page handles it itself |
npm install -g bun-browserbun-browser site update # pull community adapters
bun-browser site recommend # see which adapters match your browsing habits
bun-browser site zhihu/hot # goIf you use OpenClaw, bun-browser runs directly through OpenClaw's built-in browser — no Chrome extension or daemon required:
bun-browser site reddit/hot --openclaw
bun-browser site xueqiu/hot-stock 5 --openclaw --jq '.items[] | {name, changePercent}'Skill on ClawHub: bun-browser-openclaw
{
"mcpServers": {
"bun-browser": {
"command": "npx",
"args": ["-y", "bun-browser", "--mcp"]
}
}
}Community-driven via claw-bun-mcp. One JS file per command.
| Category | Platforms | Commands |
|---|---|---|
| Search | Google, Baidu, Bing, DuckDuckGo, Sogou WeChat | search |
| Social | Twitter/X, Reddit, Weibo, Xiaohongshu, Jike, LinkedIn, Hupu | search, feed, thread, user, notifications, hot |
| News | BBC, Reuters, 36kr, Toutiao, Eastmoney | headlines, search, newsflash, hot |
| Dev | GitHub, StackOverflow, HackerNews, CSDN, cnblogs, V2EX, Dev.to, npm, PyPI, arXiv | search, issues, repo, top, thread, package |
| Video | YouTube, Bilibili | search, video, transcript, popular, comments, feed |
| Entertainment | Douban, IMDb, Genius, Qidian | movie, search, top250 |
| Finance | Xueqiu, Eastmoney, Yahoo Finance | stock, hot stocks, feed, watchlist, search |
| Jobs | BOSS Zhipin, LinkedIn | search, detail, profile |
| Knowledge | Wikipedia, Zhihu, Open Library | search, summary, hot, question |
| Shopping | SMZDM | search deals |
| Tools | Youdao, GSMArena, Product Hunt, Ctrip | translate, phone specs, trending products |
bun-browser guide # full tutorialTell your AI agent: "turn XX website into a CLI". It reads the guide, reverse-engineers the API with network --with-body, writes the adapter, tests it, and submits a PR. All autonomously.
Three tiers of adapter complexity:
| Tier | Auth method | Example | Time |
|---|---|---|---|
| 1 | Cookie (fetch directly) | Reddit, GitHub, V2EX | ~1 min |
| 2 | Bearer + CSRF token | Twitter, Zhihu | ~3 min |
| 3 | Webpack injection / Pinia store | Twitter search, Xiaohongshu | ~10 min |
We tested this: 20 AI agents ran in parallel, each independently reverse-engineered a website and produced a working adapter. The marginal cost of adding a new website to the agent-accessible internet is approaching zero.
Without bun-browser, an AI agent's world is: files + terminal + a few APIs with keys.
With bun-browser: files + terminal + the entire internet.
An agent can now, in under a minute:
# Cross-platform research on any topic
bun-browser site arxiv/search "retrieval augmented generation"
bun-browser site twitter/search "RAG"
bun-browser site github search rag-framework
bun-browser site stackoverflow/search "RAG implementation"
bun-browser site zhihu/search "RAG"
bun-browser site 36kr/newsflashSix platforms, six dimensions, structured JSON. Faster and broader than any human researcher.
bun-browser open https://example.com
bun-browser snapshot -i # accessibility tree
bun-browser click @3 # click element
bun-browser fill @5 "hello" # fill input
bun-browser eval "document.title" # run JS
bun-browser fetch URL --json # authenticated fetch
bun-browser network requests --with-body --json # capture traffic
bun-browser screenshot # take screenshotAll commands support --json output, --jq <expr> for inline filtering, and --tab <id> for concurrent multi-tab operations.
bun-browser site xueqiu/hot-stock 5 --jq '.items[] | {name, changePercent}'
# {"name":"云天化","changePercent":"2.08%"}
# {"name":"东芯股份","changePercent":"-7.60%"}
bun-browser site info xueqiu/stock # view adapter args, example, domainThe daemon binds to 127.0.0.1:19824 by default. You can customize the host with --host:
bun-browser daemon --host 127.0.0.1 # IPv4 only (fix macOS IPv6 issues)
bun-browser daemon --host 0.0.0.0 # listen on all interfaces (for Tailscale / ZeroTier remote access)The daemon exposes the same capabilities over HTTP — use it from scripts, backends, or any language without shelling out to the CLI.
Base URL: http://127.0.0.1:19824 (same host/port as above)
Auth: every request needs a Bearer token from ~/.bun-browser/daemon.json:
export BUN_BROWSER_TOKEN=$(python3 -c "import json,os; print(json.load(open(os.path.expanduser('~/.bun-browser/daemon.json')))['token'])")
export BUN_BROWSER_HOST=http://127.0.0.1:19824Two API layers:
| Layer | Endpoints | What it does |
|---|---|---|
| Site adapters | GET/POST /site/* |
Run bun-browser site … commands — list, search, inspect, and execute adapters |
| Browser automation | POST /command, GET /status |
Low-level CDP commands — open, snapshot, click, fetch, network, etc. |
# List adapters
curl -s "$BUN_BROWSER_HOST/site" -H "Authorization: Bearer $BUN_BROWSER_TOKEN"
# Inspect args and example
curl -s "$BUN_BROWSER_HOST/site/info?name=zhihu/hot" \
-H "Authorization: Bearer $BUN_BROWSER_TOKEN"
# Run an adapter (same as: bun-browser site zhihu/hot)
curl -s -X POST "$BUN_BROWSER_HOST/site/run" \
-H "Authorization: Bearer $BUN_BROWSER_TOKEN" \
-H "Content-Type: application/json" \
-d '{"name":"zhihu/hot","args":{}}'Adapter names can contain slashes (e.g. grok/agent-chat). Use POST /site/run with a JSON body, or the path form POST /site/adapters/grok/agent-chat.
Successful runs return { "success": true, "data": { … }, "tab": "abc1", "seq": 42 }. Adapter errors return HTTP 422 with { "error", "hint", "action" } — same structured errors as the CLI.
# Health check
curl -s "$BUN_BROWSER_HOST/status" -H "Authorization: Bearer $BUN_BROWSER_TOKEN"
# List tabs (same as: bun-browser tab list)
curl -s -X POST "$BUN_BROWSER_HOST/command" \
-H "Authorization: Bearer $BUN_BROWSER_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id":"1","action":"tab_list"}'Request/response types are defined in packages/shared/src/protocol.ts. Commands queue until Chrome/CDP is connected; returns 503 if the browser is unavailable.
Full reference: OpenAPI spec · Grok adapter walkthrough
AI Agent (Claude Code, Codex, Cursor, etc.)
│ CLI or MCP (stdio)
▼
bun-browser CLI ──HTTP──▶ Daemon ──CDP WebSocket──▶ Your Real Browser
│
┌──────┴──────┐
│ Per-tab │
│ event cache │
│ (network, │
│ console, │
│ errors) │
└─────────────┘