Building a Browser Harness/MCP for AI Agents

A browser harness built with Playwright, MCP, and the GitHub Copilot SDK.

Setup

npm install
npx playwright install chromium

Set the Azure OpenAI environment variables:

export AZURE_OPENAI_ENDPOINT="https://..."
export AZURE_OPENAI_API_KEY="..."
export AZURE_OPENAI_API_VERSION="2025-04-01-preview"
export AZURE_OPENAI_DEPLOYMENT="gpt-5.4"

By default, the harness launches Playwright Chromium. To attach to Chrome instead, start it with a dedicated profile and export its CDP endpoint:

google-chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="$PWD/.agent-browser" \
  --no-first-run

export BROWSER_CDP_URL="http://127.0.0.1:9222"

Related MCP server: playwright-browser

Run

npm start -- "Open example.com and print the page title."

Run the live tasks:

npm run eval
npm run eval -- --task amazon

The code-execution runtime is not a security boundary. Run generated code in an isolated process or VM before using this pattern with untrusted input.