An AI agent for macOS is four different things

Where each category shines, and where it breaks

Terminal coding agents

Claude Code and Codex CLI are, for most engineers, the best agent loops available right now. They do real multi-step work: read a repo, plan, edit files, run tests, call MCP servers. If your work is code and you live in a terminal, the raw CLI is genuinely good and you may not need anything else.

The friction shows up at the edges. Close the terminal tab and the session is gone. Branching a conversation to try a second approach means copying a session id and resuming it by hand. On long runs the context window auto-compacts, and the summary it writes is not the summary you would have written. None of these are model problems. They are surface problems, and they are exactly the problems a native UI can fix without touching the loop.

Screenshot computer-use agents

OpenAI’s Operator, Anthropic’s computer use, and tools like Simular take a screenshot, send it to a vision model, and act on the coordinates the model returns. The appeal is universality: if a human can see it, the agent can in principle click it. The cost is real. A vision round trip per step is slow, the model misreads small UI, and a layout that shifts a few pixels breaks the plan. Several of these also run in a remote virtual machine, so the agent is not really operating your Mac, it is operating a copy.

Use this category when you genuinely need to drive software that exposes no accessibility information and no API. Outside that case, the screenshot tax is paid on every single action.

AI chat clients

BoltAI and Raycast AI are excellent native Mac apps, and they are often recommended in “AI agent” threads where they do not belong. They put ChatGPT, Claude, Gemini, and local models behind one fast window. That is useful. It is also not an agent: they generate text and code for you to copy, they do not click, type, or run anything on your machine. If what you want is fast multi-model chat, pick one of these and skip the agent question entirely.

Native-UI desktop agents

This is the category that actually takes actions on your Mac inside a native app: fazm, Dottie, EnConvo, and a few others. Most of them build their own agent loop. fazm took the opposite bet, and that bet is the rest of this page.

Where fazm sits: a native surface, not a new loop

fazm is a native Swift and SwiftUI macOS app. Underneath, it does not run a homegrown agent. It runs an ACP bridge that speaks to two adapters: the Claude Code adapter and the Codex adapter. You can check the exact versions yourself. They are dependencies in the public repo.

// acp-bridge/package.json  (excerpt)
"dependencies": {
  "@playwright/mcp": "0.0.73",
  "@agentclientprotocol/claude-agent-acp": "0.29.2",
  "@zed-industries/codex-acp": "0.12.0",
  "ws": "^8.20.0",
  "zod": "^4.0.0"
}

That choice has consequences you can feel. Because the loop is the real Claude Code, your usage hits your own Claude Pro or Max plan, and any improvement upstream lands in fazm without a rewrite. Codex is a swappable backend you select per chat, so a single window can run whichever loop fits the task. And because the loop is untouched, fazm is free to fix the three terminal frustrations purely at the surface layer.

Persistent sessions: every open chat window is written to a UserDefaults registry and restored on the next launch, so a Mac restart does not cost you a conversation. Forking: a chat has a fork button that issues a real session/fork call, opening a new window with the full prior context while the original is left alive on disk and reachable from Conversation History. And the loop reads the macOS accessibility tree and screen context rather than screenshots, so the same agent reaches past the terminal into your browser, native Mac apps, and Google Workspace. Voice input is there too: hold a hotkey and talk instead of typing.

Claims you can check before you trust any of this

Anything that can click around inside your logged-in apps should be inspectable. Every item below is greppable in the public fazm tree. If one of them is wrong, the page is wrong.

How to actually pick one

The honest version of a recommendation admits where it loses. Here is when each category wins.

• Pick raw Claude Code or Codex CLI

  if your work is entirely code, you are comfortable in a terminal, and you do not mind losing sessions on restart or forking by hand. The loop is excellent and the terminal is not in your way.

• Pick a screenshot computer-use agent

  if you must automate visual software that exposes no accessibility tree and no API. Accept the per-step latency and the brittleness, because nothing else can reach that software.

• Pick a chat client like BoltAI or Raycast AI

  if you only want fast multi-model chat and you will copy the output yourself. You do not need an agent, and pretending you do just adds permissions you will not use.

• Pick fazm

  if you want the trusted Claude Code or Codex loop, but with sessions that survive a restart, one-click forking, voice input, and reach across your browser, native Mac apps, and Google Workspace, in an open-source app you can read before you grant it permissions.

Frequently asked questions