Perplexity made its “Portable Computer” agent available in the Windows app on 14 September 2026, built with NVIDIA for GeForce RTX PCs and RTX PRO Workstations with 24GB or more of VRAM. The distinctive part isn’t the feature set — it can read and synthesise codebases, logs and documents, run multi-step workflows, generate and test code — it’s where it runs. The model, agent harness, conversation and trajectory all live on the user’s machine by default, with web search or a stronger cloud model invoked only when the user explicitly allows it.
Why this matters beyond one product launch
Search interest in “local AI agent” and “on-device AI coding” has been climbing alongside a steady stream of stories about AI coding tools leaking credentials, retaining prompts, or shipping vulnerabilities agents can exploit. Portable Computer is a direct response to that: private data never leaves the device unless the user chooses to escalate, and on-device work doesn’t consume cloud credits. It’s a similar move to what GitHub Copilot faced criticism over recently when a Claude model in its picker turned out to retain data by default — the market is now visibly splitting between cloud-first agents optimised for capability and local-first agents optimised for control.
So what
For regulated or IP-sensitive work — healthcare data, proprietary algorithms, anything you wouldn’t want leaving your infrastructure — a local-first agent option is no longer a hypothetical, it’s shipping hardware-gated but real. It’s a good prompt to ask any development partner directly where your code and data actually go when their AI tooling runs, not just what the tooling can do. We build with that question front and centre — see how we handle sensitive builds on our healthcare software development page, or get in touch if data residency is a hard requirement for your project.