Hugging Face Open Computer Agent runs a Linux VM in a browser tab. A real Mac agent reads the accessibility tree of your actual running apps.

It is a Hugging Face Space hosted at huggingface.co/spaces/smolagents/computer-agent. Under the hood it runs Qwen2-VL-72B as the vision-language model, the smolagents framework for the agent loop, and an E2B Desktop sandbox image that boots a small Linux VM with Firefox and a file manager inside. The Space renders the VM in an iframe in your browser tab. The agent reads screenshots of the VM and generates mouse and keyboard events against the VM.

Because there is no Mac in the loop. The 'computer' the agent controls is a Linux virtual machine running in a data center, exposed to you through a browser iframe. Your real apps, your real Finder, your real logged-in sessions, your real Keychain, and your real filesystem all live outside the sandbox boundary. The sandbox is the entire point of the design, it isolates the agent from every real resource.

Fazm runs as a signed macOS app on your own Mac. It ships a native ARM64 binary called mcp-server-macos-use (21,149,408 bytes) inside Fazm.app/Contents/MacOS, and registers it as an MCP server in acp-bridge/src/index.ts at line 63 (the path) and lines 1056-1064 (the spawn). The binary uses the macOS accessibility API (AXUIElement) to walk the UI tree of any running Mac app. Because the agent reads structured accessibility data rather than pixel screenshots, clicks are bound to labelled elements with resolved coordinates, not guessed from images.

The mcp-server-macos-use binary calls AXUIElementCreateApplication() on a target app's PID and walks the window hierarchy via AXUIElementCopyAttributeValue with attribute keys like kAXFocusedWindowAttribute, kAXChildrenAttribute, and kAXRoleAttribute. It writes the result to /tmp/macos-use/<timestamp>_<action>.txt as rows like `[AXButton] "Send" x:842 y:712 w:68 h:32 visible`. The agent picks a row by role and label, passes the row identifier to click_and_traverse, and the tool clicks at the centroid. No vision model guesses coordinates.

Yes, for disambiguation. Each macos-use tool call also returns a PNG screenshot at /tmp/macos-use/<timestamp>_<action>.png. When the accessibility tree is ambiguous (two buttons with the same label, a canvas-based UI like Figma, a game), the agent can read the screenshot to decide which element to target. But the baseline is the tree, and the tree is why reliability holds up across app updates that change pixels but not semantic labels.

You can run smolagents locally on your Mac, and you can spin up E2B Desktop on your Mac. But the architecture is unchanged: the agent still drives a Linux VM, not macOS. To drive macOS you would need a macOS-accessibility tool analogous to mcp-server-macos-use, and smolagents does not ship one. This is the specific gap Fazm fills.

The HF Open Computer Agent has a Firefox inside a Linux VM with a blank cookie jar. To act on any logged-in site, the agent has to log in from zero, which fails on SSO, WebAuthn, and TOTP-to-your-phone. Fazm bundles Playwright MCP running in --extension mode against a Chrome Web Store extension (Playwright MCP Bridge, ID mmlmfjhmonkocbjadbfplnigmagldckm) that attaches to your real Chrome on the Mac. Every site you are already logged into is live on turn one, and Cloudflare Turnstile and hCaptcha pass against your real browser fingerprint. The code that wires this up is in acp-bridge/src/index.ts lines 1029-1031.

Fazm is a consumer Mac app. You install it from fazm.ai as a signed, notarized .app bundle. You can inspect the bundled MCP binaries (mcp-server-macos-use, the whatsapp MCP, google-workspace MCP, the Node acp-bridge) with standard macOS tools like `codesign -dvvv`, `otool -L`, and `strings`, and you can read the bundled acp-bridge JavaScript directly in Fazm.app/Contents/Resources/acp-bridge. You do not need to trust a hosted service; the process runs locally on your Mac.

Hugging Face Open Computer Agent shines as a demo of vision-language-driven control on an arbitrary Linux desktop, especially for researchers benchmarking VLMs. Fazm shines for day-to-day Mac work: read your inbox, schedule a meeting, summarize the active Slack, fill a form in your actual Stripe Dashboard, rename files in Finder, push a git branch. The first is a research vehicle, the second is a consumer tool.

Nothing. Fazm's BUILTIN_MCP_NAMES (acp-bridge/src/index.ts line 1266) defines five built-in MCP servers, and the app accepts additional user-configured MCP servers at runtime. You could attach a smolagents-backed MCP server that fans out to an E2B sandbox, and use it from Fazm as just another tool. The architectures are complementary: Fazm is the native-Mac lane, the sandbox is the 'run this script on a throwaway Linux box' lane.

The Fazm app source tree lives on the author's Mac at /Users/matthewdi/fazm. The specific citations: acp-bridge/src/index.ts lines 63 (binary path), 1056-1064 (MCP spawn), 1266 (BUILTIN_MCP_NAMES); Desktop/Sources/AppState.swift lines 439, 441, 470, 472 (AXUIElement calls), lines 487-503 (CGEvent tap probe); Desktop/Sources/Chat/ChatPrompts.swift lines 56-61 (tool routing prompt). File sizes: /Users/matthewdi/fazm/build/Fazm Dev.app/Contents/MacOS/mcp-server-macos-use reports 21,149,408 bytes via `ls -la`.

Yes for Electron, because Chromium exposes a full accessibility tree through macOS's AX* APIs when accessibility is requested. This means Slack, Notion, Linear, VS Code, Discord, ChatGPT desktop, and Claude desktop are all readable by mcp-server-macos-use as plain accessibility trees. Native AppKit apps (Finder, Settings, Mail, Notes) work out of the box. The only tricky case is canvas or GPU-drawn apps (Figma, games), where the tree is thin and the screenshot fallback does the heavy lifting.