SilentBrowse

Stealth browser for AI agents. 95% fewer tokens. Zero detection.

npm install silentbrowse

The Problem

AI agents waste 50,000+ tokens every time they look at a web page. They send the entire DOM to the LLM, even though only a few interactive elements matter.

I ran 41 AI agents on cron jobs. My GPT Pro $200 plan ran out in 3 days.

The Solution

SilentBrowse gives agents an accessibility-tree-based view of any web page in 200-300 tokens instead of 50,000.

Wikipedia full DOM:  23,174 tokens
SilentBrowse terse:     233 tokens  (99.0% reduction)

It also passes bot detection. 20/20 on bot.sannysoft.com. Bypasses Cloudflare challenges.

Quick Start

import { launch, snapshot, click, navigate } from 'silentbrowse'

const browser = await launch({ profile: 'my-agent' })
const page = browser.pages()[0]

// Navigate
await navigate(page, 'https://news.ycombinator.com')

// Snapshot — 200 tokens instead of 10,000
const snap = await snapshot(page, {
  interactive: true,
  format: 'terse',
  maxNodes: 30,
})

console.log(snap.content)
// @e1 [link] "Hacker News"
// @e2 [link] "new"
// @e3 [link] "Show HN: Something cool"
// @e4 [link] "142 comments"
// ...

// Click by ref
await click(page, '@e3')

await browser.close()

Snapshot Formats

Terse (recommended for AI agents)

Minimal tokens. One line per element.

const snap = await snapshot(page, {
  interactive: true,
  format: 'terse',
  maxNodes: 30,
  maxTextLength: 60,
})
// snap.content = '@e1 [link] "Home"\n@e2 [button] "Sign in"\n...'
// snap.tokens = 200

JSON (for programmatic use)

const snap = await snapshot(page, {
  interactive: true,
  compact: true,
})
// snap.content = [{ role: 'link', text: 'Home', ref: '@e1' }, ...]

Tree (hierarchical structure)

const snap = await snapshot(page, {
  compact: false,
})
// snap.content = { role: 'document', children: [{ role: 'navigation', children: [...] }] }

Stealth

SilentBrowse applies 9 CDP-level patches automatically:

1. navigator.webdriver removal
2. navigator.plugins spoofing (5 fake plugins)
3. window.chrome.runtime normalization
4. navigator.deviceMemory patch (headless fix)
5. navigator.hardwareConcurrency minimum 8
6. WebGL vendor/renderer normalization
7. Permissions API patch
8. Real Chrome User-Agent (no "HeadlessChrome")
9. --disable-blink-features=AutomationControlled

Test Results

| Test | Result | |------|--------| | bot.sannysoft.com | 20/20 passed | | nowsecure.nl (Cloudflare) | Bypassed |

Token Benchmarks

| Site | Full DOM | Terse (30 nodes) | Reduction | |------|----------|-----------------|-----------| | Hacker News | 9,684 | 202 | 97.9% | | Reddit | 14,688 | 291 | 98.0% | | X (Twitter) | 1,031 | 173 | 83.2% | | Google Search | 4,011 | 208 | 94.8% | | Wikipedia | 23,174 | 233 | 99.0% |

CLI

# Start browser
silentbrowse launch --profile default

# In another terminal
silentbrowse open https://example.com
silentbrowse snap -i              # interactive elements
silentbrowse click @e1
silentbrowse type @e3 "Hello"
silentbrowse scroll down
silentbrowse close

Plugin Interface (for agent frameworks)

Drop-in replacement for browser plugins:

import { SilentBrowsePlugin } from 'silentbrowse'

const plugin = new SilentBrowsePlugin('my-profile')
await plugin.init()

await plugin.handle({ tool: 'browser.navigate', parameters: { url: 'https://x.com' } })
const snap = await plugin.handle({ tool: 'browser.snapshot', parameters: { format: 'terse' } })
await plugin.handle({ tool: 'browser.click', parameters: { ref: '@e1' } })

// Track token usage
console.log(plugin.getTokenUsage())
// { total: 215, calls: 3, breakdown: { 'browser.snapshot': 200, ... } }

API Reference

launch(options?)

| Option | Type | Default | Description | |--------|------|---------|-------------| | profile | string | 'default' | Profile directory for cookies/sessions | | proxy | string | - | Proxy URL | | headless | boolean | 'new' | 'new' | 'new' = Chrome new headless, false = headful | | executablePath | string | auto-detect | Chrome path | | cdpPort | number | - | CDP debugging port |

snapshot(page, options?)

| Option | Type | Default | Description | |--------|------|---------|-------------| | interactive | boolean | false | Only interactive elements | | format | 'json' | 'terse' | 'json' | Output format | | compact | boolean | true | Flat array vs tree | | maxNodes | number | unlimited | Limit output nodes | | maxTextLength | number | 80 | Truncate text | | selector | string | - | CSS selector to scope | | maxDepth | number | unlimited | Tree depth limit |

Actions

await click(page, '@e1')           // or CSS selector
await type(page, '@e3', 'text')
await select(page, '@e5', 'value')
await scroll(page, 'down')
await navigate(page, 'https://...')

Requirements

• Node.js >= 18
• Chrome/Chromium installed
• playwright-core (bundled)

License

MIT