rensan-browser

AI-powered browser for agents — navigate, click, type, read, see images, capture requests.

Built on Playwright + Claude. Designed for AI agents that need to interact with websites and intercept their API traffic.

Install

npm install rensan-browser playwright
npx playwright install chromium

Quick start

import { RensanBrowser } from 'rensan-browser'

const browser = new RensanBrowser({ apiKey: 'sk-ant-...' })

const session = await browser.open('https://example.com')

// Read text from the page
const headline = await session.read('the main headline')

// Click anything — AI finds it
await session.click('the Accept cookies button')

// Fill any input — AI finds it
await session.fill('Buenos Aires', 'the city search input')

// Take a screenshot (base64 JPEG)
const img = await session.screenshot()

// Capture JSON requests fired by the page
const requests = await session.capture(async () => {
  await session.fill('BUE', 'origin airport')
  await session.click('search flights button')
})
// requests = [{ url, data, score }, ...]

await session.close()
await browser.close()

API

new RensanBrowser(config)

| Option | Type | Default | Description | |---|---|---|---| | apiKey | string | required | Anthropic API key | | model | string | claude-haiku-4-5 | Vision model for AI decisions | | headless | boolean | true | Run browser headless | | timeout | number | 25000 | Default navigation timeout (ms) |

Browser methods

browser.open(url, options?)           // open session and navigate
browser.newSession()                  // open blank session
browser.openMany(urls[], options?)    // multiple sessions in parallel
browser.run(async (session) => ...)   // auto-closing session
browser.close()                       // close browser

Session methods

// Navigation
session.goto(url, options?)

// Interaction — AI finds the element via vision
session.click(goal)                   // 'the login button'
session.fill(value, goal, options?)   // ('BUE', 'origin airport input')
session.type(text, options?)          // type at current focus
session.moveTo(goal | { x, y })       // move cursor

// Reading
session.read(goal?)                   // 'the product price' or all text
session.screenshot()                  // base64 JPEG

// Scrolling
session.scroll('down' | 'up', options?)

// Request capture
session.capture(async () => { ... })  // returns CapturedRequest[]
session.startCapture()
session.stopCapture()
session.capturedRequests

// Lifecycle
session.wait(ms)
session.close()

Parallel sessions

const [s1, s2, s3] = await browser.openMany([
  'https://skyscanner.com',
  'https://kayak.com',
  'https://despegar.com',
])

const results = await Promise.all([
  s1.capture(async () => s1.fill('BUE', 'origin')),
  s2.capture(async () => s2.fill('BUE', 'origin')),
  s3.capture(async () => s3.fill('BUE', 'origin')),
])

Request scoring

Captured requests are automatically ranked by relevance — first-party API calls score higher than tracking, ads, and analytics. No hardcoded domain lists.

const requests = await session.capture(async () => {
  await session.goto('https://news-site.com')
})
// requests[0] is the most likely real data API

How AI interaction works

1. Layer 1 — Playwright semantic (free): tries getByRole('searchbox'), getByRole('textbox') and common patterns. Works for most standard inputs.
2. Layer 2 — Claude Haiku vision (~$0.001): takes a screenshot, sends it to Haiku, gets back a CSS selector. Works for any page layout.

Total cost per interaction: ~$0.001 with Haiku.

License

MIT