aibrowser 1.1.1

AI-powered browser automation

AI Browser

Let Claude Code / Cursor / Windsurf AI Agents operate the browser autonomously — stable, cheap, and automation-ready. Not a browser bot. This is the bridge layer between AI Agents and the browser.

---

Demo

Natural Language Driven

> Search GitHub for the most starred Python repos
⠋ Navigating to GitHub...
✓ Navigate to https://github.com
✓ Fill search: Python
✓ Click search button
✓ Extract top 10 repo names and star counts

Task complete:
  1. public-apis/public-apis        ⭐ 382k
  2. donnemartin/system-design-primer ⭐ 310k
  3. AUTOMATIC1111/stable-diffusion  ⭐ 280k
  ...

Benchmark Results

4 tasks × 3 runs each, measured on real pages with Chrome. Full data in benchmark_results.json.

Accuracy: 100% (12/12 runs successful)

Token efficiency:

Task	DOM tokens sent to LLM
GitHub	656
DuckDuckGo	720
Wikipedia	3,408
Hacker News	4,998

Cost per 100 tasks (3 LLM calls each):

Model	Cost
deepseek-chat	¥1.24
gpt-4o	$3.11
claude-sonnet-4-6	$4.00
Playbook replay	¥0 (zero LLM calls)

vs Browser Use

	Browser Use	AI Browser
Sent to LLM	Raw accessibility tree (~5,000–77,000 tokens)	DomRouteMap (~200–5,000 tokens)
Token reduction	—	17x lower
Repeated tasks	LLM re-interprets every time	Playbook replay, zero LLM cost
Search tasks	Opens browser → may trigger CAPTCHA	HTTP search (DuckDuckGo/Bing), no browser

---

Quick Start

1. Install

pip install aibrowser

2. Start

./scripts/start.sh --install

3. Connect to Claude Code / Cursor

Add to your MCP config:

{
  "mcpServers": {
    "aibrowser": {
      "command": "bash",
      "args": ["-c", "exec python3 python_bridge/mcp_server.py"]
    }
  }
}

4. Use

In Claude Code, just say:

> What's the top story on Hacker News today?
> Compare iPhone 16 prices on Amazon vs eBay
> Log into my company dashboard and pull this week's report

Or control manually:

goto("https://example.com")
snapshot()       # structured page view
click("@3")      # stable element index
fill("@7", "hello")
screenshot()

---

Core Features

DomRouteMap — Stable Selectors, Every Time

No fragile CSS selectors or XPaths. Pages are compressed into semantic route keys like button[Login] or textbox[Username]. LLM interaction is stable, tokens reduced ~20x.

Playbook — Learn Once, Replay Forever

Every successful action chain is auto-saved as a Playbook. Replay on the same site next time — zero LLM calls, zero cost.

# First run: LLM explores the path
aib agent "Log into Wikipedia and search for AI"
# → 5 steps, 5 LLM calls

# Second run: plays from Playbook
aib agent "Log into Wikipedia and search for AI"
# → 0 LLM calls, instant

MoE — Mixture of Experts

Complex requests are decomposed into sub-tasks assigned to specialized experts:

• Information Expert: Research without opening a browser (HTTP search + crawl + extract)
• Browser Expert: Stable DOM interaction
• Site Planner: Plan workflows using saved site profiles

Human-in-the-Loop Assistance

Behavioral simulation and interaction prompts for common human-verification scenarios, helping users complete legitimate web operations. Users are responsible for ensuring their actions comply with target websites' terms of service.

Responsible Use

AI Browser is a general-purpose automation tool designed for legitimate use cases such as:

• Testing your own websites and applications (multi-tab concurrency simulates multi-user scenarios)
• Automating your own workflows (form filling, data entry, report generation)
• Accessibility and usability research
• Compliance monitoring of publicly available information

You are solely responsible for ensuring your use complies with applicable laws and the terms of service of any websites you interact with. This tool must not be used to bypass security measures, conduct unauthorized access, or scrape data in violation of applicable laws.

Privacy Protection Mode

Optimized browser compatibility configuration to reduce environment differences in automated testing scenarios, supporting stable long-session operations.

---

Commands

Navigation:     goto, back, forward, reload
Reading:        text, html, snapshot, url, title, links, cookies, storage
Interaction:    click, fill, type, press, scroll, hover, select, upload
Inspection:     screenshot, inspect, get_attrs, is_element
Tabs:           tabs, tab, newtab, closetab, switch_tab
Sessions:       session-list, session-create, session-destroy, save_session, load_session
Agent:          agent, agent-stop, agent-status, agent-team, interrupt, history
Playbook:       playbook-list, playbook-save, playbook-replay, playbook-delete
Advanced:       js, chain, frame, state, cdp, download, privacy, handoff, shell

---

License

AI Browser is open-source under the MIT License.

Features

• 90+ browser commands
• DomRouteMap (stable selectors)
• Playbook (saved workflows)
• Multi-tab concurrency
• MoE task decomposition
• Knowledge base (semantic search)
• Task scheduler (Cron)
• Multi-tenant RBAC
• Audit log
• Unlimited agent runs
• HTTP search (no browser needed)

Activate:

aib start

---

Configuration

Variable	Description
AIBRIDGE_PORT	Server port (default: 18936)
AIBRIDGE_AUTH_TOKEN	API auth token
LLM_PROVIDER	LLM provider (openai, ollama)
LLM_MODEL	LLM model name
LLM_API_KEY	LLM API key

Security tip: Store sensitive credentials like LLM_API_KEY in your system keychain (macOS Keychain / Windows Credential Manager / Linux Secret Service) rather than plaintext config files or env scripts.

---

Tech Stack

Component	Technology
Language	Python 3.11+
Browser engine	browser-use >= 0.12.6
Browser	System Chrome/Chromium
HTTP server	aiohttp
MCP SDK	FastMCP (stdio/SSE)
LLM providers	OpenAI, Ollama, DeepSeek, GPT-4o
Vector store	ChromaDB
CDP	pydoll-python, native CDP

---

License

AI Browser is open-source under the MIT License.

This product includes third-party open-source components. See NOTICE for details.

---

🇨🇳 中文简介

AI Browser 让 Claude Code / Cursor / Windsurf 的 AI Agent 自主操作浏览器 —— 稳定、便宜、自动化就绪。

• DomRouteMap：稳定选择器，Token 降低 20 倍
• Playbook：学会就回放，零 LLM 成本
• 多 tab 并发：同一浏览器同时操作多个标签页
• 免费可用：核心功能不收费，按需升级

安装：pip install aibrowser