A browser automation tool where the browser is one tool of five, and none of them click by screenshot

Two things. First, the browser is not the whole product. Fazm registers five built-in MCP servers at runtime (fazm_tools, playwright, macos-use, whatsapp, google-workspace) and the browser is one of them, not the top of the hierarchy. You can watch this directly in acp-bridge/src/index.ts at line 1266, where BUILTIN_MCP_NAMES is literally new Set with those five names. Second, the browser leg does not operate on screenshots. Playwright is launched with --image-responses omit (acp-bridge/src/index.ts:1033), so the agent reads a YAML accessibility snapshot of the page and clicks by structural reference, not by pixel coordinates.

Those are libraries. You import a package, write code, run it in CI or from a terminal. They do exactly what you programmed, and they only drive a browser. Fazm is a signed, notarized macOS app you download and double-click. You type in English, a local AI agent decides which of the five MCP tools to call, and one of those tools is Playwright under the hood. The difference is not about raw capability on a single tab. It is that the tab is one surface in a larger plane and you are not writing code to cross it.

Browser-Use style agents take a screenshot of the page, ask a vision LLM to label the buttons, and click by pixel coordinates. That works until the page re-renders, scrolls, or the model mislabels a control. Fazm asks Playwright and macos-use for the accessibility tree of the current target, gets back a YAML snapshot with numbered refs (ref=e1, ref=e2), and asks the language model to pick a ref. No OCR, no vision step. On the Mac side the same approach uses AXUIElementCreateApplication and AXUIElementCopyAttributeValue against whatever app is frontmost, which you can see in Desktop/Sources/AppState.swift:439.

Every macOS app (including Chrome, Safari, Finder, Mail, Notes, Settings, VSCode, and Catalyst apps like WhatsApp) publishes a typed tree of UI elements to the system accessibility API. Each element has a role (button, text field, menu item), a label, bounds, and children. That tree is the same data a screen reader uses. Fazm uses it as the canonical ground truth for what is on screen, because it is structured, it does not change when a user scrolls, and it maps to a stable click target. Screenshots are lossy and nondeterministic. The tree is neither.

No. There is no CLI, no package to install, no YAML config, no test runner. You download a signed Mac app, it walks you through granting Accessibility and Screen Recording permissions, and you talk to it. Under the hood it is running Claude Agent SDK with five bundled MCP servers, but you do not see any of that unless you go read the source.

They are hardcoded in acp-bridge/src/index.ts at line 1266 as BUILTIN_MCP_NAMES = new Set(['fazm_tools', 'playwright', 'macos-use', 'whatsapp', 'google-workspace']). fazm_tools is Fazm's own helper API (database queries, screenshots for context, permissions, the ai-browser-profile). playwright is the browser leg. macos-use is a native bundled binary at Contents/MacOS/mcp-server-macos-use that drives any AX-compliant Mac app. whatsapp is a dedicated native MCP for the WhatsApp Catalyst app. google-workspace is a Python MCP for Gmail, Drive, and Calendar APIs. They are peers. The agent picks whichever one fits the sentence you typed.

Because the tasks that hurt most are the ones that cross apps. Pull a receipt from Gmail, upload it to a vendor portal in Chrome, rename the file in Finder, log the amount in a Numbers sheet. Any browser-only tool stops being useful the moment the task leaves the tab, and you are back to copy-pasting. A tool that treats the browser as one peer automates the whole chain in one request.

Yes, explicitly. In acp-bridge/src/index.ts on line 1033 the Playwright MCP is started with the flags --output-mode file --image-responses omit --output-dir /tmp/playwright-mcp. The YAML snapshot is written to a file on disk that the agent can reference, and inline base64 screenshots are dropped before they reach the model context. This is a deliberate architecture choice, not a side effect. Context stays clean and the model picks moves by ref=eN instead of guessing at pixels.

Both modes are supported. If PLAYWRIGHT_USE_EXTENSION is set to true (see acp-bridge/src/index.ts around line 1029), the Playwright MCP attaches to the Chrome you are already signed into via the Chrome Web Store extension 'Playwright MCP Bridge'. If not, Playwright launches its own headless Chromium. Real-Chrome mode is the upgrade: your cookies, SSO sessions, and browser fingerprint come along. But even fresh-Chromium mode still benefits from the structured-snapshot approach.

Inside Desktop/Sources/AppState.swift there is a function called testAccessibilityPermission (line 433) that calls AXUIElementCreateApplication on whatever app is frontmost, then reads kAXFocusedWindowAttribute back with AXUIElementCopyAttributeValue. If that round trip fails with .cannotComplete, the code does a secondary check against Finder (lines 468 to 485) to distinguish a real permission problem from an app that does not implement AX. If Finder also fails, it falls back to a CGEvent tap probe that checks the live TCC database, bypassing the per-process cache that stales on macOS 26. That self-test is why Accessibility as a primitive is trustworthy in practice.

Yes. Right after the five built-ins are registered, acp-bridge/src/index.ts reads ~/.fazm/mcp-servers.json and appends any entries with the same shape Claude Code uses: name, command, args, env, enabled. If you write a custom MCP for, say, Linear or Jira, Fazm picks it up on next launch with no code change and the agent gets a new tool immediately.

Because the user surface is the Mac app and English sentences, not a library import. The target user is somebody who wants their Mac to automate a thing, not somebody who wants to write a scraper. The fact that there is Playwright inside is an implementation detail. You would not describe Apple Shortcuts as 'a workflow engine wrapping AppleScript, Automator, and Focus'. Same framing here.