QA: Browser Use — Mastra

Mastra-only, real-LLM demo. The agent drives a real LOCAL headless browser (Playwright Chromium — no Browserbase / hosted-browser API key) via the browse_web tool and summarizes the results back into the CopilotKit chat.

Runtime requirement (read first)

• playwright is a package dependency, but the browser binary is not bundled. Install it once before running the demo:

  bashCopy

  npx playwright install chromium
  

• The demo needs a real OPENAI_API_KEY (the agent runs a live LLM) and live network access (it fetches real pages).

• The demo's Docker image must run npx playwright install chromium (and the matching OS libraries) for the browse to work in a deployed container. If the binary is missing, browse_web returns a structured error and the agent relays it — the run does not crash, but no results render.

Why there is no aimock replay / D6 fixture

Browser navigation is non-deterministic: the top Hacker News stories and any live page's contents change on every request, so there is nothing stable to record-and-replay under aimock. This cell is therefore surfaced as a Mastra-only, real-LLM demo, not a D6 aimock cell. The e2e spec (tests/e2e/browser-use.spec.ts) is a lightweight smoke test (page loads, pills render, input enabled) and does NOT drive a real browse.

Test Steps (manual / real-LLM)

• npx playwright install chromium (once)
• Navigate to /demos/browser-use
• Verify the chat surface renders with both suggestion pills: "Show me the top Hacker News stories" and "Summarize the CopilotKit homepage"
• Click "Show me the top Hacker News stories"
  • While browsing, a results card shows a "browsing…" loading state
  • After the browse, the card lists several Hacker News stories with titles (links) and point counts
  • The agent writes a short text summary mentioning a few of the stories
• Click "Summarize the CopilotKit homepage"
  • The card shows a single page-read result (title + source host) and a short text excerpt
  • The agent summarizes what the page is about in 2-3 sentences

Expected Results

• Custom per-tool renderer (useRenderTool on browse_web) renders a BrowseResultsCard with a loading state, a results list, and — if the local browser could not launch — a clear error banner (data-testid="browse-error").
• The agent never fabricates stories or page contents; it summarizes only what the tool returned.