>_ SNS-MyAgent

Browser Automation

Purpose

Let the agent drive a real browser - navigate, click, fill forms, extract content, take screenshots - for sites and web apps that need more than a fetch.

How it works

  • Built on Puppeteer (puppeteer-core + @puppeteer/browsers) with stealth scripts (src/puppeteer/*.txt) to reduce bot detection.
  • src/tools/browser/ manages the lifecycle: launch.ts (chromium launch), tab-supervisor.ts / tab-worker.ts (per-tab worker processes), tab-protocol.ts, cmux (connection multiplexing), aria/ (ARIA tree for screen-reader-style access), readable.ts (content extraction), render.ts.
  • Two modes: headless and visible; /browser toggles between them.

Configuration

browser:
  enabled: true        # master toggle (schema key browser.enabled)
  headless: true       # true = headless, false = visible window
  cmux: ...            # connection multiplexing option
  screenshotDir: ...   # where screenshots are saved

Real example

> open example.com and screenshot the hero section
> (agent launches chromium, navigates, captures docs/screenshots/...)
> /browser            # switch headless <-> visible

Screenshot

Browser mode toggle

/browser reports the active mode (headless or visible) in the TUI.

Failure behavior

  • Without a Chromium install, launch fails with a clear error; the tool surfaces it instead of crashing the session.
  • Tab workers are supervised - a crashed tab is restarted or reported rather than killing the whole browser.

Limitations

  • Requires a Chromium binary; first launch may need npx puppeteer browsers install chrome.
  • Heavier than plain HTTP fetch; use for interactions, not bulk scraping.
  • Stealth helpers reduce detection but are not guaranteed undetectable.