MCP & browser agents Updated Nov 2026 · v1.7 · Saves 6.5 hrs · Intermediate

Run stealth Browser-Use agents over persistent Chrome profiles without bot blocks

Connect Browser-Use and Playwright to your existing authenticated Chrome profile over CDP port 9222 so AI agents complete multi-step web forms and extractions without triggering CAPTCHAs.

ARCHITECTURE PIPELINE

Why default setups fall short

Fresh headless Chromium instances get flagged immediately by anti-bot walls because they lack real cookies, GPU fingerprints, and session history. The most reliable workflow shared by automation engineers on X launches a dedicated local Chrome user-data directory with remote debugging enabled on localhost:9222, then attaches Browser-Use over Chrome DevTools Protocol (CDP) with a strict Pydantic output schema.

Prerequisites before starting

Python 3.11+ with browser-use and playwright installed
Google Chrome or Chromium installed locally
An OpenAI, Anthropic, or Gemini API key for the agent controller

Step-by-step implementation

STEP 01

1. Launch a dedicated persistent Chrome profile on localhost:9222

Never run automated agents inside your personal password-manager profile. Create an isolated ~/.chrome-agent-profile directory, log in once manually to the target SaaS dashboards, and expose CDP strictly on 127.0.0.1:9222.

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="$HOME/.chrome-agent-profile" \
  --no-first-run

STEP 02

2. Attach Browser-Use to the live CDP session instead of spawning Chromium

Configure BrowserConfig(cdp_url='http://127.0.0.1:9222') so Browser-Use reuses the warm browser tabs, cookies, and TLS fingerprint without resetting session storage.

browser = Browser(config=BrowserConfig(
    cdp_url="http://127.0.0.1:9222",
    disable_security=False
))

STEP 03

3. Enforce a Pydantic Controller output schema and step cap

Always pass a Pydantic model to Controller(output_model=...) and set max_steps=18 so the agent stops immediately once structured data is captured and never loops indefinitely on unexpected modals.

result = await agent.run(max_steps=18)
parsed = ExtractionResult.model_validate_json(result.final_result())

STEP 04

4. Restrict navigation to an explicit domain allowlist

Prevent prompt injection from third-party web pages by configuring allowed_domains in BrowserContextConfig so the browser agent cannot navigate to external URLs.