Perplexity launches Portable Computer for Windows, bringing its local-first AI agent to PCs equipped with qualifying Nvidia RTX graphics hardware. The software can plan and execute multistep work on the device, use local files and tools, and call cloud models only when the user permits it.

 

The release expands an agent platform that began on Nvidia’s DGX Spark and later reached Linux systems. Windows support gives Perplexity access to a much larger professional PC market, but the initial hardware threshold keeps the product aimed at developers, researchers and power users rather than ordinary laptop owners.

 

The Windows release includes four central requirements and capabilities:

  • An Nvidia GeForce RTX or RTX Pro GPU with at least 24GB of VRAM.
  • A Perplexity Pro or Max subscription.
  • Local agent execution, model downloads and MCP integrations.
  • Scheduled tasks that can run while the user is away.

 

Related Research

 

Perplexity Launches Portable Computer on Nvidia RTX PCs

Portable Computer is the local version of Perplexity Computer, the company’s cloud-based agent for complex work. Instead of only answering a prompt, the system can break a goal into smaller steps, assign work to subagents, use connected tools and assemble a final result.

 

On a compatible Windows machine, the orchestrator, subagent model and agent harness can run locally. Perplexity says users can work with files and connected applications without sending the task to its servers. Local execution also avoids consuming the cloud credits attached to Computer tasks.

 

The system is not locked to the device. When a job requires stronger frontier reasoning, Portable Computer can ask permission to route a step to a cloud model. This hybrid design gives users a choice between local privacy and the additional capability of remote models instead of silently moving every request online.

 

The 24GB VRAM Requirement Narrows Windows Access

Perplexity and Nvidia require at least 24GB of graphics memory for on-device inference. That makes GPU memory, rather than the age of a PC, the primary entry barrier. Compatible consumer cards include the GeForce RTX 3090, RTX 4090 and RTX 5090, while several RTX Pro models also meet the threshold.

 

Many mainstream gaming laptops and desktops ship with less memory, even when their GPUs are otherwise powerful. The first version therefore targets machines already configured for local model development, high-end content work or technical workloads. Owners of qualifying desktops can use existing hardware rather than buying a dedicated DGX Spark system.

 

Perplexity provides one-click local model downloads inside the Windows app. Nvidia highlighted Qwen 3.8 27B as one option, while Perplexity also offers its post-trained PPLX 27B model. The 27-billion-parameter scale helps explain why the product needs substantially more memory than typical consumer AI features.

 

Local MCP and Scheduling Expand the Agent Workflow

The Windows version adds support for local Model Context Protocol servers. MCP gives the agent a standardized route to desktop tools and private services that a user controls. Perplexity says customers can connect their own integrations instead of relying only on the company’s built-in connector catalog.

 

Nvidia listed Microsoft Word, Google Drive, Gmail, Slack and GitHub among supported connections. The exact privacy boundary depends on the connector and task: local files can stay on the PC, while using an online service naturally requires communication with that service.

 

Scheduled tasks move the product beyond an assistant that waits for a prompt. A user can assign recurring work and leave the agent running on the Windows machine. That could support periodic document processing, repository checks or research updates, provided the PC remains available and the connected tools retain permission.

 

Local AI Changes the Cost and Privacy Tradeoff

Cloud agents are convenient because providers manage the models and compute, but they also meter usage and require data to cross an external service. Portable Computer moves more of that cost to hardware the user already owns and can keep sensitive material within a locally controlled execution path.

 

The approach does not make every workflow private by default. Cloud escalation, web research and third-party connectors can still transmit information. The meaningful change is that the user has a local option for selected steps and explicit control when Perplexity proposes sending work to a frontier model.

 

Running agents locally also creates new operational responsibilities. Users must secure the machine, review tool permissions and understand what a scheduled task can change. An agent with access to local files and desktop applications can be useful precisely because it has privileges that require careful limits.

 

Windows Becomes a Test for Local AI Agents

Portable Computer arrives as Nvidia and PC makers promote computers built for on-device AI. Nvidia expects Windows systems using its RTX Spark platform from Lenovo and Acer, while Perplexity’s release gives owners of existing high-memory RTX hardware a way to test a local agent now.

 

The immediate audience is small compared with the wider Windows market, but it is technically influential. Developers and professional users are more likely to experiment with local models, MCP servers and unattended workflows, creating feedback that can shape less demanding versions for future consumer systems.

 

Perplexity’s broader wager is that AI agents will not remain entirely cloud services. If the Windows version proves reliable, local execution could become a standard option for work involving proprietary documents, custom software or predictable recurring tasks—especially as PCs gain more memory and stronger inference hardware.