Perplexity launched Portable Computer on August 25, a fully local version of its agentic “Computer” platform that runs the model, planner, tools, and sandbox on the user’s own hardware. At launch it supports Nvidia DGX Spark and Linux machines with compatible RTX GPUs (24 GB VRAM minimum). Windows support is promised for September. The system starts every task on-device. Only when the local model hits a limit does it ask permission to escalate a specific step to a cloud frontier model. Local inference consumes zero token credits.
This is a meaningful shift for anyone who has been paying ongoing API bills or worrying about where client data travels. Sensitive documents, internal SOPs, customer lists, or proprietary code can stay on the machine. The agent still connects to Google Drive, Gmail, Slack, and GitHub when needed, but the heavy lifting happens locally. Early users report that routine research, document analysis, and multi-step file work feel faster and more private because nothing leaves the device unless explicitly approved.
Hardware is the current gate.
A high-end Nvidia setup is required today, which puts it out of reach for many pure software-only solopreneurs. Yet the direction is clear: serious agent workloads are moving toward a hybrid local-first model. For founders who already own or can access compatible hardware—or who handle regulated or confidential work—the combination of privacy, predictable cost, and reduced latency is compelling. Even those who cannot run it yet should watch the September Windows release and the promised expansion to more consumer GPUs.
My recommendation is: Test it if you can (if you have or can access compatible Nvidia hardware) otherwise watch the Windows rollout