Perplexity has made it possible to run its whole Computer agent on a Mac, from the model that reasons to the program that plans the steps. It starts on the machine, works offline, and calls a cloud model only when the owner allows it. Hybrid Compute in September worked the other way round: it started in the cloud and sent only the private steps to the Mac. A similar local version had already reached Windows PCs with Nvidia graphics cards, so what is new is that Mac owners get it too.
A metered tool gets used less than a tool you own
Perplexity prices Computer in credits, 100 to the dollar, and a light task uses roughly 100 of them. If a morning run that turns your mail and messages into a priority list is a light task, it costs about $1. That is $22 a month for one person who works 22 days, and $220 for a team of ten.
Nobody refuses to pay that. But the dollar makes people stop and ask whether each run is worth it, and the pause is the real cost. It is the same hesitation that made mainframe time something people planned for. The spreadsheet mattered because a manager could test a plan the moment it occurred to them, and so they tested far more plans.
A recruiter with 150 unanswered candidate messages shows what that looks like. She could have the agent read all of them, put each person's availability and salary expectation into a Numbers sheet, and draft replies. In the cloud she might run that once a week. On her own Mac she could run it every hour, and the candidates' personal details would never leave the machine. She still reads each draft before anything is sent, which is where her judgment belongs.
Memory is the new price tag
The model has to fit in the Mac's memory alongside the apps it works with, which is why Perplexity recommends 36GB. That is a recommendation, not a hard minimum, but it is more than some current Macs can even be configured with. The local models are Qwen 3.6 and a version of it that Perplexity trained for Computer.
So the question that matters when a firm buys laptops for staff who handle client files will move from how fast the machine is to how much memory it has. The local model is also smaller than the more than 15 frontier models it can call on, so the sensible pattern is the one Perplexity built. The Mac does the routine and private work, and you approve a cloud step only when a job needs deeper reasoning.
What to hand it first
The first jobs for a local agent are the ones you never automated because each run felt like spending money, and the files you have kept away from every cloud tool. List both. A machine you have already paid for makes the first set cheap to try and the second set possible.