Vercel AI SDK

ai-sdk-tool-antibrow gives an AI SDK agent four browser tools that all drive the same persistent profile. A tool that starts a fresh Chromium is fine for reading a public page and useless for anything behind a login: the session dies with the process. Here the cookies, the storage, the fingerprint and the exit IP are the same ones the agent used last time.

Install

npm install ai-sdk-tool-antibrow
export ANTI_DETECT_BROWSER_KEY=ab_live_...

ai and zod are peer dependencies - the versions your app already has. The browser engine downloads on first launch and is cached. Unlimited local profiles and one concurrent browser are free; the SDK reference has every launch option.

Use it with generateText

import { gateway, generateText, stepCountIs } from 'ai'
import { antibrowTools } from 'ai-sdk-tool-antibrow'

const browser = antibrowTools({ profile: 'research-01' })

try {
  const { text } = await generateText({
    model: gateway('openai/gpt-5-mini'),
    prompt: 'Open https://example.com and tell me the heading.',
    tools: browser.tools,
    stopWhen: stepCountIs(5),
  })
  console.log(text)
} finally {
  await browser.close()
}

Close the session in a finally: a closed browser leaves nothing behind, an abandoned one is a whole Chromium still running. The same object works with streamText and with any agent loop that takes AI SDK tools - they are ordinary tool() definitions, with nothing AntiBrow-specific about how they are called.

The four tools

antibrowTools() is the shortcut: one session, four tools, one close(). Each tool is also exported on its own, so you can take two of them or rename them - what matters is that they share a session, which is what makes them composable: a read reads whatever the last navigation opened.

import { AntibrowSession, browserGoto, browserRead } from 'ai-sdk-tool-antibrow'

// Only want two of them, or want to name them yourself:
const session = new AntibrowSession({ profile: 'research-01' })
const tools = { open: browserGoto(session), read: browserRead(session) }

Options

antibrowTools({
  profile: 'research-01',                // same name -> same fingerprint, cookies, storage
  proxy: 'http://user:pass@host:5001',   // one exit IP per profile
  temporary: true,                       // discard the profile on close
  headless: false,
  readLimit: 4000,                       // characters a read may return
  key: process.env.ANTI_DETECT_BROWSER_KEY,
})

Two agents that must not share an identity need two profile names, not two tabs. How many may run at once is your plan's concurrency limit, and a launch over it raises rather than silently queueing.

How this was tested

Two checks ship with the package (src/test.ts). First the tools run for real, against a live profile:

const browser = antibrowTools({ profile: 'ai-sdk-selfcheck', temporary: true })
await browser.tools.browserGoto.execute({ url: 'https://example.com' }, ctx)
// -> { status: 200, title: 'Example Domain', url: 'https://example.com/' }
await browser.tools.browserRead.execute({ selector: 'h1' }, ctx)
// -> { url: 'https://example.com/', truncated: false, text: 'Example Domain' }

Then the whole AI SDK tool-call path runs with a mock model standing in for the LLM, so the input schema, the tool-call dispatch and the tool result round-trip are exercised rather than assumed:

tool results -> [
 {
  "type": "tool-result",
  "toolCallId": "call-1",
  "toolName": "browserGoto",
  "input": { "url": "https://example.com" },
  "output": { "status": 200, "title": "Example Domain", "url": "https://example.com/" },
  "dynamic": false
 }
]

What it does not do

It does not promise that a given site will accept an automated session. A persistent identity removes the tells that come from starting over every run - a fresh profile, a stock automation fingerprint, your own IP - and that is all it removes. Our dated measurements, including the checks that fail, are published under reports. Driving a browser from an agent without writing code? See the MCP server; using LangChain instead? See the LangChain guide.