Perplexity Personal Computer for Mac, the practical guide as of May 2026

What changed on May 7, 2026

The April 16 launch was the kind of release that mostly mattered to people already paying $200 a month for Perplexity Max. The May 7 release is the one that touches the rest of the user base: the Pro tier ($20 per month) and Enterprise both get the agent, and the entire Mac app got rebuilt around it. The old Mac app is being deprecated in the weeks after, so this is not a parallel option, it is a forced migration on a soft schedule.

The four weeks between the two launches matter for one reason: every concrete piece of feedback you read about Personal Computer dated before May 7 was written by a Max subscriber on a high-tier plan, often with hardware Perplexity supplied or recommended. The Pro experience is the one most people will actually evaluate, and the volume of public review for that started accumulating only this week.

The five questions every Mac user is asking

Below, the same set of questions you would actually ask a friend who installed it before you. Read in order or jump around.

1. Which Mac do I need

Anything on macOS 14 Sonoma or later. Perplexity is explicit that the heavy lifting runs on their servers, so an entry-level M1 MacBook Air handles Personal Computer at the same shape of latency as a maxed-out M3 Pro. The local agent is doing screen reads, click injection, file I/O, and the network round-trip to Perplexity. None of those are CPU-bound on modern Apple Silicon.

The recommended setup if you want the agent to keep working while you are away from the keyboard is a Mac mini left running. That is what the launch coverage on MacRumors led with on April 16: “Turning a Mac mini Into an Always-On AI Agent”. The mini is the cheapest Apple Silicon Mac with no battery to manage, no lid to close, and no thermal throttling story to think about. A used M1 mini is roughly $400 today, which is two months of Pro plus the device.

2. Pro at $20, or Max at $200

After May 7, both tiers can run Personal Computer. The agent itself is the same product. The difference is upstream: Max bundles “over 20 frontier models” that the orchestrator can route between for different shapes of task; Pro gets a more conservative model rotation. Max also tends to ship first on new capabilities (it had four weeks of head start on this one). For most small-business users, Pro is the right floor; Max becomes interesting only if you already use the heavy model rotation on the chat side.

If your only reason to be on Max was Personal Computer access, that reason has now collapsed into the cheaper tier. There is a real argument for downgrading.

3. What apps does it actually reach

The marketing language is “works across any Mac app” plus Comet for the web. The named integrations called out at the May 7 launch were Microsoft Teams, Excel, and 1Password, on top of the over-400 connector library Perplexity maintains. The list below is the surface area as of this week.

What the surface area does not tell you is the depth in each app. “Works across any Mac app” is a coverage claim, not a capability claim, and the gap between “the agent can read the focused window of Notes” and “the agent can complete a real Notes-driven workflow without a human nudging it” is the entire game. The named connector list (Teams, Excel, 1Password) is the part with the most publicly demonstrated reliability; the “any Mac app” part is best treated as a moving target you will have to test against your own apps.

4. Where does the work actually run

Perplexity calls the architecture “hybrid”. In practice that means: the Mac app handles the keyboard hotkey, file system access, the local sandbox, and the connection to Comet. Everything Perplexity calls “intense”, which is the orchestrator that picks tools, the model inference, and any longer-running planning, runs on Perplexity infrastructure. The Mac app is the eyes and hands. The brain is hosted.

That is the same line every mainstream computer-use product currently draws, and it is reasonable. It is also worth being explicit about, because it shapes what you can and cannot say about the agent in a privacy review or a SOC2 questionnaire. “The agent runs on my Mac” is true for the click injection layer and false for the orchestration layer, and which half of that sentence matters depends on what you are buying it for.

5. Why is it not in the Mac App Store

The May 7 build is a direct download from perplexity.ai, not an App Store distribution. TechCrunch’s coverage flags this explicitly. The likely reason is App Store sandboxing: an agent that needs to read other apps’ window state, send synthesized keystrokes, and access files across app boundaries cannot fit cleanly inside the App Store sandbox rules. A direct-download Developer ID build can request the same Accessibility, Screen Recording, and Automation permissions any third-party Mac utility (Alfred, Raycast, Bartender) asks for, then operate within those grants.

Practically, this means you grant three sets of System Settings permissions on first launch. There is no “just install from the App Store and it works” path. That is true for every shipping computer-use agent on macOS today; it is not specific to Perplexity.

How to set it up, end to end

Five concrete steps, in order. Most of the friction is in step four (the permissions panes), which is the same friction every Mac user experiences with any tool that touches Accessibility.

The architectural detail no launch article surfaces

Read every launch article cover to cover and you will not find a sentence on which macOS API Personal Computer uses to read the focused window of Mail, click the Send button in Slack, or fill a row in Numbers. The candidates are: Apple’s public Accessibility framework (AXUIElement), the older Apple Events / AppleScript bridge, screen recording plus pixel-coordinate click injection, or some hybrid. Each has a different reliability and permission profile. A reader who cares enough to ask is left guessing.

That gap is a market signal. Closed binaries get to be vague about implementation; open ones do not. The closest like-for-like macOS computer-use agent that publishes its implementation is Fazm. The agent loop runs through a Node bridge at acp-bridge/src/index.ts that spawns a documented set of MCP subprocesses: fazm_tools (first-party tool surface), playwright (browser automation), macos-use (native macOS app control via the Accessibility API), whatsapp (Catalyst app control), and google-workspace (Apps Script + OAuth).

That set is wired in the same file at lines 1535 through 1758, with user-defined servers from ~/.fazm/mcp-servers.json and inherited Claude Code servers from ~/.claude.json layered on top by name. Anyone can read it. None of this is hypothetical or a slide-deck claim; it is the runtime configuration of a shipping product, on github.com/m13v/fazm.

That is not an argument that Personal Computer is wrong to be closed. It is an argument that, if the implementation question matters to you, there is exactly one place on Mac to look at the answer in code, and Perplexity’s product is not that place.

When Personal Computer is the right answer

To be fair to it: if you already pay for Perplexity Pro for search and the AI overviews, the agent is bundled at no extra cost, and the Comet browser integration is a real advantage for web-heavy workflows. The 400+ pre-wired connectors are work that someone has already done for you. The Mac mini always-on pattern is genuinely well-supported. If your work is mostly web research orchestration plus a handful of named SaaS integrations, and you do not mind the orchestrator running on someone else’s servers, the trade is reasonable. It is a polished consumer product and the team has shipped fast.

The shape that points the other way is when the agent has to live inside your apps for hours of unattended work, when you want the lines of code that touch your microphone or your file system to be readable, or when you need a Mac-native experience that does not break if a vendor changes its mind about pricing. Different shapes, same general problem. Pick the one that matches the constraint you actually have.