Automation of a web browser, done by reading the accessibility tree, not the pixels

It usually means one of three things. One, a dev framework that writes code against a headless Chromium (Playwright, Puppeteer, Selenium). Two, a cloud service that rents you a headless browser over an API (Browserless, Firecrawl, Cloudflare Browser Run). Three, an AI browser replacement (Perplexity Comet, Chrome Auto Browse) that runs in a sandbox separate from your Mac. Fazm is none of those. It is a consumer macOS app that attaches a local AI agent to the Chrome window you are already signed into, reads the page as a YAML accessibility snapshot rather than a screenshot, and the same agent continues into Mail, Finder, Calendar, and WhatsApp once the browser leg is done. The tool routing for all of this lives in Desktop/Sources/Chat/ChatPrompts.swift lines 56 to 63.

Snapshot, almost always. When Fazm's agent calls the playwright browser_snapshot tool, the Node bridge writes a YAML file under .playwright-mcp/ that contains the accessibility tree for the page: every role, label, value, and an integer reference like [ref=e17]. The agent then calls browser_click with ref: 'e17' — no CSS selector, no XY coordinate, no pixel-reading. Screenshots are a fallback reserved for 'when you need visual confirmation' because, in the words of the system prompt, 'it costs extra tokens.' The choice matters because screenshots lose the button's role, a page redesign breaks a selector but a ref computed against the live tree survives it, and accessibility trees see background content that a screenshot cannot crop into.

Consumer. You install a signed, notarized Mac app. The first time you run a flow that needs the browser, Fazm opens a 880x520 setup window that walks you through installing Google Chrome (if missing), adding the Playwright MCP Bridge extension from the Chrome Web Store (ID mmlmfjhmonkocbjadbfplnigmagldckm), clicking the puzzle piece icon to copy the token, and pasting it. Four numbered steps on the left, animated GIFs on the right. No CLI. No npm install. No YAML to hand-edit. If you want a Python SDK you want playwright-python; if you want 'ask the AI to do the thing in your own Chrome,' you want Fazm.

After any navigate, click, or type action, the agent gets a compact YAML file on disk at .playwright-mcp/<timestamp>.yml. Each interactive element gets a bracketed ref — [ref=e3], [ref=e17] — that stays stable across the same page load. The agent uses those refs for browser_click, browser_type, browser_fill_form, browser_select_option. Only when a human would need to visually verify something (chart contents, a captcha picture, a layout bug) does it call browser_take_screenshot, and even then the system prompt tells it to prefer capture_screenshot (desktop-wide) over browser_take_screenshot (viewport only), because the agent's real work is on the whole Mac, not inside the browser tab.

Your real Chrome. The setup checks for /Applications/Google Chrome.app explicitly, and a 2-second polling timer waits until it appears if you have not installed it. The bridge then sets the env var PLAYWRIGHT_USE_EXTENSION=true and passes your auth token as PLAYWRIGHT_MCP_EXTENSION_TOKEN (wired at Desktop/Sources/Chat/ACPBridge.swift lines 368 to 377). From that point on, every playwright call runs inside the Chrome you have been using for months, with your cookies, your saved logins, your installed extensions, and the browser fingerprint that a site has already seen. There is a non-extension fallback that launches its own Chromium if you skip the setup, but you lose the session benefit.

No. The routing prompt has an explicit tab-hygiene rule (ChatPrompts.swift line 63): before navigating to a site, the agent calls browser_tabs action 'list' to check if you already have it open. If a match is found by domain, it switches via browser_tabs action 'select' instead of opening a new tab. If the current tab matches, it reuses that one. After finishing a task, tabs the agent opened are closed with browser_tabs action 'close'. The only time a new tab is opened is when the user explicitly asks for one. This is the difference between 'an agent sharing your browser' and 'an agent living in its own browser.'

Constantly, if the job is a real workflow. Paying a bill ends in Mail. Scheduling something ends in Calendar. Following up ends in WhatsApp or iMessage. Downloading a file ends in Finder. The problem with every cloud-headless 'browser automation' product is that the moment the workflow crosses that boundary, the automation stops and you take over manually. Fazm registers four automation targets in the same bridge: playwright (Chrome), macos-use (Finder, Settings, Mail, Notes, any AX-compliant app via the native mcp-server-macos-use binary), whatsapp (the native Catalyst app), and google-workspace (Gmail, Drive, Calendar APIs). One agent, one message, four surfaces.

Familiar, but inverted. The Playwright API surface is there — browser_navigate, browser_snapshot, browser_click, browser_fill_form, browser_type, browser_take_screenshot, browser_tabs — but you never write against it yourself. You describe the result you want, the model picks which tools to call, reads each response, and composes the next step. There is no test file, no BeforeAll, no Page Object, no CI. If the page layout shifts in a way a scripted selector would not survive, the model re-reads the new snapshot and picks the new ref. If the job was 'reply to that thread,' it just runs.

They are real products, they are just solving a different problem. Perplexity Comet and Chrome Auto Browse are agentic browsers. They replace your browser with theirs, run tasks in their container, and show you a summary. Fazm does not replace anything: it drives the Chrome you already use, keeps your tabs, keeps your bookmarks, and also drives your native apps. If your job is read-and-summarize the web, a new AI browser is fine. If your job is 'do a multi-app thing on my Mac that happens to involve the web,' a fresh browser is the wrong primitive.

The model is. Fazm's default is the Claude API, but the ACP bridge reads a user-configurable ANTHROPIC_BASE_URL (Desktop/Sources/Chat/ACPBridge.swift lines 379 to 382) from the customApiEndpoint UserDefault. Point it at a local OpenAI-compatible shim in front of Ollama, LM Studio, or vLLM and the model call is local too. The automation layer — playwright, macos-use, whatsapp — is all local either way: the browser extension talks to the local Node bridge over a loopback token; the native mcp-server-macos-use binary ships inside Fazm.app/Contents/MacOS and never phones home.