Your FBI agent is fired. Meet the AI agent you actually hired.

No. 'The FBI agent assigned to my computer' is a TikTok-era meme: a friendly fiction in which an exhausted federal employee has to watch your search history all day. The Bureau is not literally watching your webcam, and the only way they could is via a court order plus targeted exploitation, not a default subscription. Quora and Clario have written this answer at length. The interesting question for 2026 is the inverse: there genuinely is going to be an agent on your computer, in the form of an LLM-driven assistant, and the question is whether you choose it, see it, and audit it. Fazm is one answer to that question.

Three concrete differences, all visible in the source. First, when Fazm is driving the browser the page itself wears a halo: acp-bridge/browser-overlay-init.js injects four pulsing wing gradients (#fazm-w-top through -right) and a centered pill that reads 'Browser controlled by Fazm · Feel free to switch tabs or use other apps' (line 58), at z-index 2147483647 so nothing can hide it. Second, the way Fazm 'sees' your Mac is by reading the macOS Accessibility tree (the same one VoiceOver uses) via the bundled mcp-server-macos-use (registered at acp-bridge/src/index.ts line 1056), and by reading labelled DOM nodes with [ref=eN] tokens via Playwright; not by taking screenshots and asking a vision model to guess pixel coordinates. Third, the onboarding flow is required to send 'Everything is open-source at github.com/mediar-ai/fazm — your data is stored locally and you own it all.' (ChatPrompts.swift line 308) before it can ask for any permission. Other agents do not have to do any of that.

An accessibility-API read returns a structured text tree: every interactive element has a role (Button, TextField, Group), an optional label ('Send message'), a position rect, and sometimes an editable value. That tree is the same one screen readers use to speak the screen aloud. A screenshot is a 2D array of RGB pixels; everything else (what is a button, what does it say, where is the cursor) has to be guessed by a vision model. Reading the tree is faster, more reliable across UI redesigns, and auditable in a way pixels are not. You can save the tree to a log file and read it back. You cannot meaningfully audit what a vision model 'saw' in a screenshot.

Three guardrails that surveillance does not have. One, it requires a macOS permission you actively granted in System Settings, which you can revoke at any time. The OS itself enforces the gate, not Fazm. Two, the only process reading the tree is one you can terminate by quitting the app, and the Python MCP server speaks stdio rather than listening on a port (verifiable with lsof). Three, the data goes to local conversation history and to whatever AI provider you have configured for your turn; the agent's instructions to be transparent about what it is doing are baked into the system prompt at session start (Desktop/Sources/Chat/ChatPrompts.swift). Surveillance is invisible, persistent, network-bound, and deny-by-default for the user. Fazm is the inverse on every axis.

Because anything else would feel like the meme. The halo is not branding for its own sake; it is a hard requirement of consensual automation. acp-bridge/browser-overlay-init.js paints four edge-pinned radial gradients (#fazm-w-top, #fazm-w-bottom, #fazm-w-left, #fazm-w-right) on a 4 to 5 second pulse and a centered pill. The element id is 'fazm-overlay', the canvas inside it has z-index 2147483647 with !important, so a hostile or buggy site cannot occlude it. If you cannot see the halo on a page, Fazm is not driving that page. That is the contract.

z-index 2147483647 (the maximum 32-bit signed integer, applied to both #fazm-canvas2 on line 23 and #fazm-pill3 on line 43 of browser-overlay-init.js) plus !important on every layout property in the pill block. A site that wanted to draw on top would need to match the same value, and CSS specificity rules with !important make that very hard to do silently. The overlay is also re-injected on every page load and DOMContentLoaded event by the patched Playwright extensionContextFactory.js (scripts/patch-playwright-overlay.cjs), so a single-page-app navigation cannot strip it.

Local first. The macos-use server speaks stdio (no port), the bundled Python MCP for Google Workspace also speaks stdio (no port), conversations and the knowledge graph live on disk in your user data directory. When the agent makes an LLM call, that of course goes to whichever model provider you have configured. There is no analytics-driven 'send a copy of every screenshot to a vendor server' pipeline; there is no copy of every screenshot, because the agent reads the accessibility tree instead.

Yes. The onboarding script literally requires the agent to say so before requesting any permissions: ChatPrompts.swift line 308 instructs the agent to send 'Everything is open-source at github.com/mediar-ai/fazm — your data is stored locally and you own it all.' as a trust-building message before the first sensitive prompt. You can read the overlay file, the MCP server registrations, the build script that bundles the Python interpreter, and the Swift floating bar code. None of it is obfuscated.

Yes. Two surfaces. First, the Fazm chat panel logs every tool call: each browser_snapshot, each macOS accessibility query, each Python invocation appears as an inline tool call card you can expand. Second, the on-page overlay tells you the browser is under control; absence of overlay means absence of browser control. For native Mac app reads, the agent's chat reply contains the exact tool calls and their text-form arguments, because the macos-use server returns text. You cannot get the same audit trail out of any screenshot-based agent.

Quit the app from the menu bar. The bridge process and the five MCP servers exit with the app (acp-bridge/src/index.ts line 1266 enumerates them). If you only want to disconnect the browser, open chrome://extensions and revoke the Playwright MCP Bridge token, which kills the CDP attach instantly. There is no covert resident process that survives quitting; that is the whole point of building this on plain macOS app architecture rather than a kernel extension.

Because the meme is the right cultural anchor for explaining what AI agents on the desktop should look like. The TikTok 'FBI agent assigned to my computer' is funny because everyone agrees, intuitively, that an invisible non-consenting watcher is bad. Fazm is the easiest possible product to position against that intuition: a Mac app you actively install, that asks for permissions one at a time, that paints itself onto every page it touches, that reads your UI through APIs you can audit, and that ships its source code. If your computer is going to have an agent on it (and it is) the meme is the test for what good looks like.