An AI agent that works at your desk — and keeps the folder at home

On Tuesday, August 25, Perplexity — the company behind the answer-engine with citations — shipped Portable Computer. Built with Nvidia, it is a version of the firm’s agent system that runs on hardware you own, or just bought: Nvidia’s DGX Spark, a desk-sized AI box, and Linux PCs with a hefty Nvidia graphics card.

This is not “another chat.” Chat is a question and an answer. An agent, in the sense the industry is using now, is a system that takes a job and keeps going: opens a folder, reads many files, crunches numbers, writes a report, sometimes posts the result to Slack. That burns far more than a conversation. In the cloud, each step can hit the bill. On Portable Computer, if the work stays on the device, the company says it consumes no billing credits.

The detail reporters kept from Monday’s demos was visual. Nate, a Perplexity vice president, had the system read a folder of tax and investment documents — the kind of paper many people will not upload to someone else’s server. A local 27-billion-parameter Qwen model went file by file. The number that usually climbs in the corner of the screen, the cloud-credit tally, “is just parked at zero,” he said, “because all of this is happening on the device.”

There is a middle path. The system starts on the computer. If a step needs a stronger model in the cloud, it asks before sending. Before that send, a check looks for personal data in what would leave the machine, and the person sees the outgoing slice. The remote model returns text guidance; it does not touch local files. In one demo, after analyzing a spreadsheet on the device, the agent posted the result to a Slack channel. Google Drive, Gmail, and GitHub are on the connector list too.

Why is Nvidia in the room? For two years the company has sold the world giant data centers. Now Nader, a director there, says local AI “reached an inflection point”: it left the hobby of shrinking tiny models and became a practical tool, thanks to larger, useful open models. For Nvidia, software that actually uses a DGX Spark helps sell a box that, people at the briefing admitted, has been easier to buy than to put to work.

Who can use it today is not “anyone with a laptop.” At launch: Pro, Max, and enterprise subscribers, on Linux. Windows is due in September. The graphics bar is 24 GB of video memory — roughly an RTX 3090 or newer. Mac, home to many people who tinker with AI at home, is missing. Nate was blunt: the focus right now is Nvidia hardware.

The company published a paper arguing that a small model and the agent’s scaffolding have to be designed together. Scaffolds built for giant cloud models want huge context and dozens of tools at once; the local model buckles. Perplexity shortened the prompt, kept few tools loaded, and pulls the rest only when needed. If the operating-system sandbox is down, the harness shuts itself off — rather than running commands with the user’s full permissions, as some open kits do by default.

The prettiest scores come from Perplexity’s own internal test, which it says it will open. That asterisk belongs in the story. Even in that slice, the local model still trails a frontier cloud model on the hardest tasks. The middle path — local first, cloud only with a yes — closes part of the gap for less money, on the company’s own math.

The economic argument is easy to translate. A chatbot turn is short. An agent that spends hours in a folder is a different bill. In the cloud, that bill grows. At the desk, the extra cost of each generated word falls toward zero, and the sensitive file does not have to cross the country. For a law office, a hospital, or a finance shop, that weighs as much as the price.

None of this is everyone’s computer. It is a launch for people who already pay the subscription and already have (or will buy) an expensive card. What the news marks, for a reader who only wants to know what is going on, is a different meter: after two years of measuring AI in gigawatts and cloud invoices, a well-known company is trying to sell the opposite — the heavy work at your desk, the counter at zero, the folder that never travels.