Perplexity brings Portable Computer to Windows RTX PCs

Perplexity has released its local-first agent, Portable Computer, for Windows PCs with NVIDIA RTX GPUs. The core agent runtime, including its orchestrator, planner, models, tool routing, task queue and local search index, runs on the user’s machine. The Windows release also adds local Model Context Protocol servers for desktop-app access and a scheduler for recurring jobs.

Portable Computer debuted in August on NVIDIA DGX Spark before expanding to Linux PCs with compatible RTX hardware. Windows brings the same architecture to a broader base of consumer and professional computers, giving developers a way to run long agent workflows without paying for every local inference step or sending every file to a hosted model.

What stays on the PC

Perplexity packages Portable Computer with models configured for its agent runtime and optimized for NVIDIA GPUs. The model picker includes Qwen3.5 27B and PPLX 27B, Perplexity’s post-trained variant. The company says Nemotron 3.5 Lightning, NVIDIA’s 30B open model, will join the picker.

• Agent control: A deterministic software orchestrator coordinates planning, tool selection, sandboxed execution, task queues, scheduling, local indexing and persistent state.
• Desktop-app access: Local MCP servers expose application tools and data to the model. MCP is an open protocol that standardizes how AI systems call external tools and retrieve context.
• Cloud-service connectors: Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub integrations route through the local orchestrator.
• Recurring jobs: The scheduler can launch tasks while the user is away from the keyboard, provided the PC remains available.

Local MCP execution keeps planning and tool calls on the device. Connectors to hosted services still communicate with those providers, so the data path depends on the applications involved in each workflow.

Cloud help requires approval

Portable Computer can request help from a cloud model when a task exceeds the local model’s reasoning capacity. Before sending anything, the orchestrator selects relevant context, checks it with a personally identifiable information classifier and displays the proposed material for user approval.

The remote model receives the approved context and returns text advice. It has no direct access to local files, applications or tools. The local orchestrator retains control and continues the task after the response arrives.

Small models reshape the agent loop

A 27B local model has less reasoning capacity than a large hosted model, while agent workflows can accumulate extensive histories of plans, tool calls and results. Perplexity redesigned its harness around those constraints:

• The core system prompt is smaller.
• Task-specific skills load only when required.
• Compact command-line interfaces replace some verbose MCP tool definitions.
• Older trajectory data is compressed as a task grows.

Perplexity reports that Qwen3.5 27B exposes a context window of about 260,000 tokens but shows declining agent performance beyond roughly 100,000 tokens. Portable Computer therefore manages the useful context available to the agent instead of filling the model’s entire nominal window.

Who can run it

Requirement	Details
GPU	NVIDIA RTX GPU with at least 24 GB of VRAM, such as an RTX 5090, a suitable RTX PRO card or another supported equivalent
Operating system	Windows through the existing Perplexity app; compatible Linux PCs and DGX Spark are also supported
Subscription	Perplexity Pro at $20 per month or Max at $200 per month
Installation	The Windows app downloads the required local model weights and configures the runtime
Cloud access	Optional, with user approval before selected context leaves the device

Local tokens change the bill

Agent workloads consume more tokens than short chat sessions because one request may trigger repeated planning, parsing, search, coding and verification steps. Perplexity says some of its evaluation trajectories reach hundreds of thousands of tokens for a single task.

Running those loops locally shifts the expense from per-token API charges to hardware, electricity and GPU depreciation. Perplexity says on-device work consumes zero account credits. That claim covers local execution; approved cloud escalations fall outside it.

Best suited to repeated private work

Portable Computer fits high-volume, privacy-sensitive tasks grounded in local files and applications, including:

• Summarizing folders of PDFs in batches
• Running module-by-module code migrations across a repository
• Triaging pull requests on a schedule
• Reconciling spreadsheets against internal files

With local MCP integration, the agent can operate desktop applications without sending their contents to Perplexity when the workflow remains on the PC. Tasks involving Gmail, Slack, GitHub or other hosted services still exchange data with those providers, and cloud-advisor calls transmit the context a user approves.

Workflows dominated by frontier-model reasoning, extensive browser automation or current web data will rely more heavily on hosted services, which can make a cloud agent simpler. Recurring jobs over local files and desktop tools gain more from Portable Computer’s privacy controls and zero-credit local inference, provided the user can justify a subscription and an RTX GPU with at least 24 GB of VRAM.