Anthropic ships three distinct Claude products that touch a Mac in 2026: Claude Desktop (the native app), Mac control (Computer Use), and Claude in Chrome (the browser extension). They sit at different layers of the system. They have different capabilities. Knowing how they relate is the difference between using Claude effectively on a Mac and stumbling around wondering why one feature does not work in the same place another one does.
This piece is a clean conceptual map. By the end you will know what each layer does, what each cannot do, and how Safari fits into the picture.
Layer 1: Claude Desktop (the native Mac app)
Claude Desktop is the native macOS application you install from claude.ai. It is the same Claude conversation interface that lives at claude.ai in your browser, packaged as a native app with system-level integrations: keyboard shortcuts, menu bar access, file drag-and-drop, screenshot capture, and a few Mac-specific UX flourishes.
Without any additional features enabled, Claude Desktop is a fast, native conversation interface that:
- Lives in your menu bar for instant access
- Accepts file attachments via drag-and-drop
- Stores conversation history locally
- Provides keyboard shortcuts for new conversations and quick capture
- Renders code, tables, and structured output cleanly
It does not, by itself, do anything to other applications. It does not see Safari. It does not see your terminal. It does not see what is on your screen. It is a chat app.
Layer 2: Mac control via Computer Use
Mac control is the May 13, 2026 release that enabled Computer Use on Claude Desktop. It is a feature inside Claude Desktop, not a separate product, but it is meaningful enough that thinking of it as a distinct layer helps.
When Mac control is enabled (you grant accessibility and screen recording permissions in System Settings), Claude can:
- Take screenshots of your screen
- See what applications are running
- Interact with applications at OS level based on tier (click in full-tier apps, click but not type in click-tier apps, only screenshot in read-tier apps)
- Run shell commands via Terminal
- Read clipboard contents
This is the layer that introduces the tier system. Browsers are read-only. Terminals and IDEs are click-only. Most other applications are full-tier.
Without Mac control enabled, Claude Desktop has none of these capabilities. It is a regular chat app. Mac control is an opt-in capability you enable when you grant the relevant system permissions.
Layer 3: Claude in Chrome (browser extension)
Claude in Chrome is a separate Anthropic product shipped as a Chrome extension. It is currently in beta and limited to a controlled cohort of Claude Max subscribers, with broader rollout expected in 2026.
Claude in Chrome:
- Sees the current Chrome tab\'s URL and page content
- Can click, type, scroll, and navigate inside Chrome
- Has its own permission and consent model separate from Computer Use
- Operates only inside Chromium-based browsers (Chrome, Edge, Brave, Opera, Arc)
- Does not work in Safari (Safari is not Chromium)
This is the layer where Anthropic puts action-tier browser work. Chrome is the platform; the extension is the permission boundary; Claude in Chrome is the product that lives on top of both.
Where Safari fits
Safari sits in an awkward position relative to all three layers:
Relative to Claude Desktop (Layer 1)
Safari has no special relationship to Claude Desktop. You can use Claude Desktop while Safari is open the same way you can use it while any other app is open. They do not talk to each other.
Relative to Mac control (Layer 2)
Safari sits at read tier. Claude can take screenshots of Safari but cannot click, type, or scroll inside it. Claude can use shell commands (via Terminal) to indirectly drive Safari with open -a Safari "URL". The OS handles the URL; Safari just receives it. Read-only tier is respected because Claude never clicks inside Safari directly.
Relative to Claude in Chrome (Layer 3)
No relationship. Claude in Chrome only operates inside Chromium-based browsers. Safari is excluded.
The third-party layer (AlliHat)
One additional product worth mapping: AlliHat, the third-party Safari extension. AlliHat is not from Anthropic. It does not sit inside any of the three Anthropic-shipped layers. It is a separate product that connects to Claude via the Anthropic API.
AlliHat:
- Lives inside Safari as a Safari App Extension
- Authenticates against your Anthropic account
- Provides sidebar UX and Agent Mode actions inside Safari
- Costs $29 per year on top of your existing Claude subscription
For users who want sidebar Claude inside Safari with action capabilities, AlliHat is the only commercial product that delivers. For users who only want read-and-analyse research, the shell workflow via Computer Use\'s read-only Safari tier is sufficient and free.
How the layers compose in practice
A typical research session on a Mac in 2026 mixes layers:
- Claude Desktop (Layer 1) is open in the menu bar, ready for conversation
- Mac control (Layer 2) is enabled, so Claude can run shell commands and screenshot
- Safari is the user\'s primary browser, sitting at read-only tier per the tier system
- For read-and-analyse tasks: curl + open -a Safari + screenshot, all driven by Mac control
- For action-driven tasks: open Chrome and use Claude in Chrome (Layer 3) for the specific URL
- Optional: AlliHat installed for sidebar Claude inside Safari with Agent Mode
The Mac power user setup runs both browsers and uses Claude Desktop with Mac control as the orchestration layer.
What is missing from the map
One thing Anthropic has not shipped that would simplify all of this: an official Safari extension. If it existed, Safari would move from read-only to action-tier under Computer Use (or alongside it), the AlliHat workaround would not need to exist, and the Mac story would be much cleaner. As of May 2026, Anthropic has explicitly said this is "not planned."
The strategic logic
The product architecture makes sense once you see it. Anthropic puts dedicated, audited extensions at the action-tier browser layer. Computer Use treats browsers as observe-only by default. This is the safest design: no accidental form submissions, no accidental purchases, no accidental account changes. Users who want action-tier browser interaction install the dedicated extension and consent to that capability explicitly.
The cost of this design is that one major browser (Safari) does not have an official action-tier path. The benefit is that no user accidentally gets action-tier behavior they did not consent to.
What this means for you
- If you use Mac and Safari, install Claude Desktop and enable Mac control. The shell-based workflow covers most research tasks.
- If you need action-driven workflows in a browser, install Chrome and Claude in Chrome alongside Safari.
- If you refuse to install Chrome and need action-driven workflows in Safari, install AlliHat.
- If you only do read-and-analyse work, the shell-based workflow is enough.
Bottom line
Three Anthropic-shipped layers (Desktop, Mac control, Claude in Chrome) plus one third-party layer (AlliHat) cover almost every Claude + Mac scenario in 2026. Safari sits awkwardly because Anthropic has not shipped a dedicated Safari product, but the workarounds are well-understood and work today.
At ScaleWise VA we use Claude Desktop + Mac control + Claude in Chrome as our standard team setup. If you want help mapping the right stack for your own workflow, book a free 30-minute call.
How this stack maps to a Shopify operations team
At ScaleWise VA, every operations specialist has the full Mac stack described above configured on their workstation:
- Claude Desktop in the menu bar for ad-hoc questions and quick research
- Mac control enabled for the shell workflow during competitor research and supplier audits
- Chrome + Claude in Chrome as the secondary browser specifically for Shopify admin work (where action-tier capabilities matter)
- Safari as the primary browser for everything else (battery life on MacBook Air, native Apple ecosystem)
Specialists rotate between Safari and Chrome about a dozen times per workday depending on whether the task is read-and-analyze (Safari + shell workflow) or action-driven (Chrome + extension). The split feels natural after a week of practice.
For clients who hire our team, this complexity sits entirely on our side, we deliver the operational outcomes without you having to think about which browser does what. Book a discovery call and we'll walk you through what our team's actual workflow looks like.