If you've evaluated MCP servers for browser automation, you've seen three credible options: Microsoft's Playwright MCP , Browserbase + Stagehand , and Tap . They look like substitutes. They aren't. Same tool slot. Three different products. Each answers a different question. The architectural axis that separates them: where does the browser actually run, and what credentials does it see? Where browser runs Logged-in / cookies Tokens per run Trust boundary Playwright MCP Local Playwright (headless or --extension ) △ --extension reuses your Chrome; headless = none Per-call (LLM at runtime) Your machine Browserbase + Stagehand Browserbase cloud ✗ credentials must be uploaded Per-call (LLM at runtime) Browserbase's cloud Tap Your own Chrome (extension) ✓ uses your live session 0 (deterministic replay) Nothing leaves the machine Tokens: the architecture, not the engineering For "scrape the top 30 stories from HN as JSON," runtime-extracting tools call an LLM each invocation.…