Most AI agents still have one big problem: they live in the cloud. Your files are sent away for processing, and every serious task can eat into your usage limits.
Perplexity thinks it
has a better idea. Its new Portable Computer runs the AI agent on your own hardware, with no token costs for work done locally.
There is a catch, though. You are going to need a pretty serious PC.
The whole AI agent runs locally
Portable Computer is a local version of Perplexity Computer, the agent platform the company launched earlier this year. Instead of sending every task to a remote server, it puts the models, agent tools, inference engine and security sandbox on your own machine.
That means it can read PDFs, work with local files, run commands and handle longer tasks without uploading everything to the cloud. It can also connect to services such as
Gmail,
Google Drive, Slack and
GitHub.
More importantly, local tasks don't use cloud credits or generate token costs.
I can see the appeal here. If you use AI all day for work, paying for every token starts to matter. Keeping private files on your own computer is also a much better option for some tasks.
It can still ask the cloud for help
Running locally doesn't mean Portable Computer is cut off from more powerful AI models. If the local model can't handle part of a task, the agent can ask for permission to send that step to a cloud model. Perplexity says the user must approve the move first. The rest of the task can stay on the PC.
That's probably the part I like most.
Small local models are getting much better, but they still can't match the best cloud models in every task. A system that stays local most of the time and asks before sending something away makes more sense than pretending the cloud isn't useful.
At launch, Portable Computer supports Qwen 3.8 27B and Perplexity's own PPLX 27B model, with Nvidia Nemotron 3.5 Lightning coming later. Users can also bring their own model and inference server.
Your normal laptop probably won't run it
Now for the bad news. Portable Computer launches on Linux and needs Nvidia hardware with at least 24GB of VRAM. It works with the Nvidia DGX Spark and compatible RTX systems running DGX OS or Ubuntu. Windows support is planned for September, while there is no macOS version yet.
So no, this isn't something most people will install on a $700 laptop today. It is also available to Perplexity Pro, Max, Enterprise Pro and Enterprise Max subscribers.
Still, I think the idea matters more than the first version. AI agents are getting access to our email, documents and other private data. Running that work on hardware you own, while using the cloud only when you allow it, feels like a much better direction.
The hardware just needs to get cheaper.