Give an OpenAI agent a browser.

pip install openai-agents playwright browsa

export BROWSA_API_KEY="agt_live_..."
export OPENAI_API_KEY="sk-..."

import asyncio
import os
from dataclasses import dataclass

from agents import Agent, RunContextWrapper, Runner, function_tool
from browsa import Client
from playwright.async_api import Page, async_playwright

BROWSA_API_KEY = os.environ["BROWSA_API_KEY"]
TASK = "Open https://example.com, follow its link, and tell me the title of the page you land on."


@dataclass
class Browser:
    page: Page


@function_tool
async def open_url(ctx: RunContextWrapper[Browser], url: str) -> str:
    """Open a URL in the browser. Returns the page title."""
    page = ctx.context.page
    await page.goto(url, wait_until="domcontentloaded")
    return f"Opened {page.url}. Title: {await page.title()}"


@function_tool
async def read_page(ctx: RunContextWrapper[Browser]) -> str:
    """Return the URL and the visible text of the current page."""
    page = ctx.context.page
    text = await page.inner_text("body")
    return f"URL: {page.url}\n\n{text[:8000]}"


@function_tool
async def click(ctx: RunContextWrapper[Browser], text: str) -> str:
    """Click the first link or button whose visible text contains `text`."""
    page = ctx.context.page
    await page.get_by_text(text).first.click(timeout=10_000)
    await page.wait_for_load_state("domcontentloaded")
    return f"Clicked. Now on {page.url}. Title: {await page.title()}"


@function_tool
async def fill(ctx: RunContextWrapper[Browser], label: str, value: str, press_enter: bool) -> str:
    """Type `value` into the input whose label or placeholder matches `label`."""
    page = ctx.context.page
    box = page.get_by_label(label).or_(page.get_by_placeholder(label)).first
    await box.fill(value, timeout=10_000)
    if press_enter:
        await box.press("Enter")
        await page.wait_for_load_state("domcontentloaded")
    return f"Filled. Now on {page.url}"


agent = Agent[Browser](
    name="Browser agent",
    instructions=(
        "You control a real Chrome browser through tools. "
        "Use read_page to see the page before you click. Answer in one or two sentences."
    ),
    tools=[open_url, read_page, click, fill],
    model="gpt-5.6-luna",
)


async def main():
    browsa = Client(api_key=BROWSA_API_KEY)
    cloud = browsa.create_agent(ttl="5m")  # a Browsa cloud browser, 1 credit
    try:
        async with async_playwright() as p:
            browser = await p.chromium.connect_over_cdp(
                cloud["cdp_http"], headers={"X-API-Key": BROWSA_API_KEY}
            )
            context = browser.contexts[0]
            page = context.pages[0] if context.pages else await context.new_page()

            result = await Runner.run(agent, TASK, context=Browser(page), max_turns=20)
            print(result.final_output)
    finally:
        browsa.destroy_agent(cloud["id"])


if __name__ == "__main__":
    asyncio.run(main())

python agent.py