Agentcore Browser

AgentCore Browser — CDP over AWS Managed Chrome

Automate a cloud-hosted Chrome 144 browser through Bedrock AgentCore. The browser runs entirely on AWS — no local Chrome, no X server, no Playwright install. You talk to it via CDP commands over a SigV4-authenticated WebSocket.

When to Use This Skill vs agent-browser

Try agent-browser first. It is faster (local execution, no network RTT) and free. Switch to this skill when:

| Situation | Why this skill wins | |-----------|-------------------| | No local Chrome / headless server | Browser runs on AWS, nothing to install | | agent-browser CLI not installed | Only needs boto3 + websockets (Python) | | Python agent pipeline (Strands, LangGraph) | Native async/await, no shell subprocess | | AWS VPC-internal sites | VPC network mode reaches private ALBs | | Need session persistence across runs | Browser Profile API saves cookies/storage | | agent-browser --headed display issues | Not applicable — always headless on AWS |

Prerequisites

pip install boto3 websockets

AWS credentials must have bedrock-agentcore:* and bedrock-agentcore-control:* permissions.

Quick Start

Step 0: Create a Browser Resource (one-time)

import boto3
control = boto3.client("bedrock-agentcore-control", region_name="us-east-1")
resp = control.create_browser(
    name="myBrowser",
    description="Experiment browser",
    networkConfiguration={"networkMode": "PUBLIC"},
)
browser_id = resp["browserId"]  # e.g. "myBrowser-O9bw0JnQyc"

Or use the helper:

from scripts.browser_client import AgentCoreBrowser
browser_id = AgentCoreBrowser.create_browser("myBrowser", region="us-east-1")

Step 1: Connect and Navigate

import asyncio
from scripts.browser_client import AgentCoreBrowser

async def main():
    browser = AgentCoreBrowser(browser_id="myBrowser-xxxxx", region="us-east-1")
    await browser.start("my_session")

    await browser.navigate("https://example.com")
    title = await browser.get_title()
    print(f"Page title: {title}")

    await browser.screenshot("/tmp/page.png")
    await browser.stop()

asyncio.run(main())

Core Workflow

Every automation follows this pattern:

1. Navigate — await browser.navigate(url)
2. Extract — await browser.evaluate(js) or await browser.get_page_text()
3. Interact — await browser.click(selector) / await browser.type_text(selector, text)
4. Capture — await browser.screenshot(path)
5. Stop — await browser.stop() (always call this to avoid charges)

API Reference

Lifecycle

| Method | Description | |--------|-------------| | AgentCoreBrowser.create_browser(name, region) | Create a browser resource (class method, one-time) | | AgentCoreBrowser.list_browsers(region) | List existing browser resources | | browser.start(name) | Start session, connect WebSocket, attach to page | | browser.stop() | Close WebSocket and stop session | | browser.destroy_browser() | Delete the browser resource |

Navigation

| Method | Description | |--------|-------------| | navigate(url, wait=2.0) | Go to URL, wait for load | | wait_for_selector(selector, timeout=10) | Poll until element exists |

Data Extraction

| Method | Description | |--------|-------------| | evaluate(expression) | Run JavaScript, return value | | get_text(selector) | Get element's innerText | | get_page_text() | Get full page text | | get_html(selector) | Get element's outerHTML | | get_url() | Current page URL | | get_title() | Current page title |

Interaction

| Method | Description | |--------|-------------| | click(selector) | Click element by CSS selector | | type_text(selector, text) | Type text char-by-char (CDP key events) | | select_option(selector, value) | Select dropdown option |

Capture

| Method | Description | |--------|-------------| | screenshot(path=None) | PNG screenshot, optionally save to file |

Patterns

Form Submission

await browser.navigate("https://httpbin.org/forms/post")
await browser.type_text('input[name="custname"]', "Test User")
await browser.type_text('input[name="custemail"]', "test@example.com")
await browser.evaluate(
    "document.querySelector('input[name=\"size\"][value=\"medium\"]').checked = true"
)
await browser.evaluate("document.querySelector('form').submit()")
await asyncio.sleep(2)
result = await browser.get_page_text()

Web Scraping (JS-rendered pages)

await browser.navigate("https://finance.naver.com/marketindex/")
rate = await browser.evaluate(
    "document.querySelector('.head_info .value')?.innerText"
)
print(f"USD/KRW: {rate}")

Wait for Dynamic Content

await browser.navigate("https://spa-app.example.com")
found = await browser.wait_for_selector(".data-table", timeout=15)
if found:
    data = await browser.evaluate("document.querySelector('.data-table').innerText")

Important Notes

Cost Management

Sessions are billed by duration. Always call browser.stop() in a finally block:

browser = AgentCoreBrowser(browser_id="...", region="us-east-1")
try:
    await browser.start("task")
    # ... work ...
finally:
    await browser.stop()

IP Restrictions

The browser runs on AWS IP ranges. Some sites block AWS IPs:

• weather.go.kr (Korean Met Administration) — returns error page
• accuweather.com — returns 403 Access Denied

Sites that work well from AWS:

• naver.com, finance.naver.com — fully functional
• httpbin.org — fully functional
• wttr.in — fully functional
• Most SaaS/enterprise tools — generally accessible

Form Input

Setting element.value = "..." via JS alone does not trigger browser input events on all form frameworks. The type_text() method uses CDP Input.dispatchKeyEvent to simulate real keystroke events, which works reliably across React, Vue, and plain HTML forms.

CDP Protocol

The automation stream speaks Chrome DevTools Protocol v1.3 (JSON-RPC over WebSocket). Any CDP command works — the methods above are convenience wrappers. For advanced use:

result = await browser._send(
    "Network.enable", session_id=browser._cdp_session_id
)

Browser Profile Persistence

Save and restore session state (cookies, localStorage) across sessions:

# Save after login
client = boto3.client("bedrock-agentcore", region_name="us-east-1")
client.save_browser_session_profile(
    profileIdentifier="my_profile",
    browserIdentifier=browser_id,
    sessionId=session_id,
)

# Reuse in next session
resp = client.start_browser_session(
    browserIdentifier=browser_id,
    name="resumed",
    profileConfiguration={"profileIdentifier": "my_profile"},
    ...
)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jesamkim
• Source: jesamkim/oh-my-skills
• License: MIT

Install and usage instructions live in the source repository linked above.