Skip to content

AshtonVaughan/agentbrowser

Repository files navigation

AgentBrowser

A browser runtime built for AI agents. Not humans.

Every existing browser automation tool — Playwright, Puppeteer, Selenium, browser-use — was built for humans first and retrofitted for agents. They speak in DOM operations. Agents think in tasks.

AgentBrowser inverts this. The browser speaks the agent's language.


The Core Difference

Every other tool:

// Agent receives 8,000 tokens of HTML noise
// Agent guesses what "#submit-btn-v2" means
// Agent issues a DOM command and hopes
await page.click('#submit-btn-v2')
await page.fill('input[name="email"]', email)

AgentBrowser:

// Agent receives 50 tokens of structured meaning
// Agent sees exactly what it can do
// Agent calls a semantic action
const state = await browser.navigate('https://example.com/login')
// state.page_type        → 'login'
// state.available_actions → ['authenticate', 'signup', 'forgot_password']
// state.key_data         → { site: 'Example', sso_available: true }

await browser.action('authenticate', { email, password })
// result.state_change.summary → 'Navigated from login to dashboard'
// result.next_available_actions → ['view_profile', 'settings', 'logout']

Features

Semantic observation — Agents never see HTML. Every page navigation returns a structured model: what type of page it is, what data it contains, what actions are available. Token cost drops ~95%.

Dynamic tool registry — The available tools change as the page changes. On a login page, authenticate() appears. On a checkout page, submit_order() appears. The agent always sees exactly what it can do — nothing more.

Site memory — The browser learns permanently. First visit to a site: full LLM analysis. Second visit: cache hit, 7x faster, zero LLM cost. Tenth visit: the LLM gets injected context of proven selectors and known page flows, producing more accurate results from the start.

Self-healing execution — CAPTCHA detection, stale selector recovery, and post-action state verification are handled silently. Agents see task outcomes, not infrastructure failures.

Bot detection bypass — Ships with playwright-extra + stealth plugin. Bypasses Cloudflare challenges, cookie consent banners (OneTrust, Cookiebot, GDPR), and signup modals automatically on every navigation. No configuration needed.

Session persistence — Sessions are first-class objects. Save, restore, and branch browser state across agent runs. Auth state survives restarts.

MCP server — Ships as a Model Context Protocol server. Any MCP-compatible agent (Claude Code, LangChain, AutoGen, custom) connects without integration work.

Parallel tasks — Declare goals, not tabs. The runtime manages contexts, isolation, and result aggregation.


Tested Sites

Results from parallel 7-site test (14s total):

Site Page Type Detected Notes
news.ycombinator.com listing Actions execute (navigate, read, submit)
en.wikipedia.org article Cookie banner auto-dismissed
github.com landing Full action set available
stackoverflow.com listing Signup modal stripped before analysis
www.bbc.com listing OneTrust consent auto-dismissed
www.theverge.com listing Cloudflare challenge bypassed
www.reddit.com blocked IP-level network block — no client-side bypass

Installation

git clone https://github.com/AshtonVaughan/agentbrowser
cd agentbrowser
npm install
npx playwright install chromium
npm run build

Set your Anthropic API key:

cp .env.example .env
# Edit .env and set ANTHROPIC_API_KEY

Usage

As a library

import { AgentBrowser } from './src/index.js'

const browser = new AgentBrowser({
  anthropic_api_key: process.env.ANTHROPIC_API_KEY,
  headless: true,
  stealth: true,
})

await browser.launch()

// Navigate — returns semantic model, not HTML
const state = await browser.navigate('https://news.ycombinator.com')
console.log(state.page_type)          // 'listing'
console.log(state.available_actions)  // ['navigate_to_new', 'navigate_to_ask', ...]
console.log(state.key_data)           // { story_count: 30, top_story: '...' }

// Execute actions by name — actually clicks/navigates
await browser.action('navigate_to_new')
// → browser navigates to https://news.ycombinator.com/newest

// Extract structured data
const data = await browser.extract({
  top_story: 'title of the #1 story',
  points: 'upvote count of top story',
  author: 'username of top story submitter',
})

// Save session (auth state, cookies)
await browser.saveSession('hn-logged-in')

// Restore later
await browser.restoreSession('hn-logged-in')

await browser.close()

Run multiple sites in parallel

const results = await browser.executor.runParallel([
  { id: 'hn',   goal: 'get top stories', url: 'https://news.ycombinator.com' },
  { id: 'bbc',  goal: 'get headlines',   url: 'https://www.bbc.com/news' },
  { id: 'gh',   goal: 'check trending',  url: 'https://github.com' },
])
// All three run concurrently in isolated browser contexts

As an MCP server

Add to your Claude Code config (~/.claude.json):

{
  "mcpServers": {
    "agentbrowser": {
      "command": "node",
      "args": ["/path/to/agentbrowser/dist/server/mcp.js"],
      "env": {
        "ANTHROPIC_API_KEY": "your-key-here",
        "AGENTBROWSER_HEADLESS": "true",
        "AGENTBROWSER_STEALTH": "true"
      }
    }
  }
}

Then in any conversation, Claude has access to:

Tool Description
navigate Go to a URL, get semantic page state
get_page_state Current page type, data, available actions
page__<action> Dynamic tools generated from current page
extract Pull structured data via schema
fill_form Fill a named form by field labels
start_session Begin a new browser session
save_session Persist cookies and auth state
restore_session Resume a saved session
run_parallel Execute multiple tasks simultaneously
get_memory_stats View accumulated site knowledge

How the Learning Works

First visit to stripe.com/login
  → Full LLM analysis (~4 seconds)
  → Stores: page model, page type, selectors

Second visit (within 30 min)
  → Cache hit (< 500ms, zero LLM cost)

Third visit (cache expired)
  → LLM gets injected context:
      "This domain has been visited 8 times.
       Known page types: login, dashboard, billing, settings
       Proven selectors:
         authenticate.email:    input[id='email']
         authenticate.password: input[type='password']
         authenticate.submit:   button[data-qa='submit']"
  → More accurate output, correct selectors from the start

After each successful action
  → Winning selector saved to selector library
  → Site profile updated with transition knowledge
  → Cache invalidated on navigation

The browser gets permanently smarter about each site. Knowledge compounds across sessions and agents.


Bot Detection & Popup Handling

AgentBrowser handles anti-bot measures automatically:

Stealth — Uses playwright-extra with the full stealth plugin suite. Patches WebGL, canvas fingerprint, navigator.webdriver, plugins, permissions, chrome.runtime, and more. Passes most bot detection checks including Cloudflare.

Consent auto-dismissal — After every navigation, automatically clicks through:

  • OneTrust / Cookiebot / Funding Choices banners
  • Wikipedia cookie prompts
  • Reddit GDPR gates
  • Stack Overflow notice bars
  • Generic "Accept all" / "Got it" / "Agree" buttons

Force removal — If no button is found, JS-removes known overlay elements (.s-modal, cookie policy overlays, full-viewport fixed backdrops) so they don't pollute the semantic analysis.

Modal stripping — Before sending HTML to the LLM, modal/consent/overlay elements are stripped from the input so they can't mislead page classification.


Architecture

Agent
  │
  │  semantic intent (not DOM commands)
  ▼
┌─────────────────────────────────────┐
│           Task Runtime              │  ← self-healing, parallel, recovery
│  ┌───────────────────────────────┐  │
│  │      Site Memory Store        │  │  ← SQLite: cache, selectors, profiles
│  └───────────────────────────────┘  │
│  ┌───────────────────────────────┐  │
│  │      Semantic Analyzer        │  │  ← Claude Haiku: page → structured model
│  └───────────────────────────────┘  │
│  ┌───────────────────────────────┐  │
│  │       Browser Engine          │  │  ← playwright-extra: stealth Chromium
│  └───────────────────────────────┘  │
└─────────────────────────────────────┘
  │
  │  HTTP
  ▼
 Web
Component File Role
Types src/types.ts SemanticPageModel, ActionDefinition, AgentSession
Browser Engine src/engine/browser.ts playwright-extra wrapper, stealth, consent dismissal, sessions
Semantic Analyzer src/semantic/analyzer.ts LLM page analysis with site context injection
Site Memory src/memory/store.ts SQLite: page cache, selector library, site profiles
Task Executor src/runtime/executor.ts Self-healing execution, learning feedback loop
MCP Server src/server/mcp.ts Dynamic tool registry over MCP protocol
Public API src/index.ts AgentBrowser convenience class

Configuration

Option Default Description
anthropic_api_key required Anthropic API key for semantic analysis
headless true Run browser headlessly
stealth true Enable anti-bot detection bypass
memory_db_path ~/.agentbrowser/memory.db SQLite database path
semantic_max_tokens 2048 Max tokens for page analysis
mcp_port 3100 MCP server port

Why Not Just Use Playwright?

Playwright is exceptional at what it does. But it was built to automate what humans do manually — and it shows. Agents using Playwright must:

  • Parse thousands of tokens of HTML noise to find the one thing they care about
  • Re-learn every site from scratch on every session
  • Translate between intent ("log in") and DOM operations ("find input, fill, click") on every action
  • Handle CAPTCHA, popups, and stale selectors themselves
  • Manage tab state manually for parallel work

AgentBrowser is not a wrapper around Playwright. It's a different abstraction layer — one where the browser's mental model matches the agent's mental model from the ground up.


License

MIT

About

Browser runtime purpose-built for AI agents — semantic tools, site memory, MCP server

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors