Perplexity opened Personal Computer in its Mac app to all users. It says the agent works across local files, native Mac apps, the web, and Perplexity secure servers. Pricing, permission boundaries, and task success rates are not disclosed. My reaction is caution, because this pushes agent failure onto the user’s actual desktop.
Desktop agents are no longer a novelty. OpenAI’s Operator focused on browser tasks, and Anthropic showed computer use with Claude 3.5 Sonnet controlling screens. Perplexity is choosing the Mac as the entry point, closer to Raycast, Spotlight, and a browser assistant. That is a stronger surface than search, but it also carries much sharper failure modes.
The Mac is not a web sandbox. If an agent reads local files, controls native apps, and calls remote servers, the permission model becomes the product. The post does not say whether there is per-task approval, folder allowlisting, sensitive-action blocking, or audit logs. Without those, Personal Computer smells like a remote intern with root-adjacent habits.
Perplexity’s core edge has been retrieval and web context, not OS automation. It won trust by attaching answers to citations and live search. Desktop work needs execution reliability, rollback, and state awareness. Filling a wrong field in a browser is annoying; editing the wrong local file is a different liability class.
I have doubts about “any Mac” as a deployment claim. AppleScript, Accessibility permissions, sandboxing, and macOS version drift create messy edge cases. Until Perplexity publishes task benchmarks or failure handling, I would not treat this as a mature agent platform. The product proof is not opening Finder; it is surviving boring workflows without quietly breaking things.