Perplexity Portable Computer Comes to Windows RTX PCs

Perplexity (@perplexity_ai) announced on September 14, 2026 that Portable Computer — its local AI agent harness — is now available on Windows PCs equipped with NVIDIA RTX GPUs. It is the third operating system for a product that launched Linux-only in August, and it lands with two new capabilities: local MCP connections and scheduled tasks. CEO Aravind Srinivas (@AravSrinivas) framed it as "unmetered local intelligence on every Windows PC running on NVIDIA hardware and Perplexity harness," and NVIDIA (@nvidia) posted supporting content alongside the launch.

That framing is worth testing against the actual hardware requirement, because it's easy to read "every Windows PC" as more inclusive than it is. explainx.ai already tracked the RTX-vs-DGX-Spark story on Linux — confirming an RTX 3090 works, not just the $4,679 DGX Spark — and that same post flagged Windows as "announced for September 2026, not yet shipped." This is that promise landing. But the hardware gate that applied on Linux applies here too, and it's more restrictive than the "unmetered on every PC" pitch suggests.

TL;DR — what people are asking

The 24GB VRAM bar, precisely

This is the detail that determines whether "available on Windows" actually means available to you. Perplexity's requirement for on-device inference is an NVIDIA RTX GPU with 24GB of VRAM or higher — not "an RTX GPU," which is a much lower bar most gaming PCs already clear, but specifically 24GB or more.

That number matters because of what it excludes:

• Laptops are largely out. Even RTX-branded gaming laptops typically top out around 16GB of VRAM on their highest-end mobile GPUs. A 24GB card is a desktop-class or workstation-class part.
• Mainstream desktops are out. Most prebuilt and budget gaming towers ship with 8GB-12GB cards. The 24GB floor points at cards like the RTX 3090, RTX 4090, or workstation RTX GPUs — the same territory explainx.ai's MacBook vs. dedicated GPU comparison already treats as the practical entry point for serious local inference, and the same class of card running frontier-scale local models at home.
• This mirrors the Linux requirement exactly. Perplexity didn't loosen the bar to reach Windows users — it's the identical 24GB floor from the DGX Spark and RTX Linux launch, just extended to a second operating system.

What actually runs locally

Portable Computer runs the harness, the agents, and the underlying models on the user's own PC. That includes working with local files and connected apps — Perplexity names Outlook, Gmail, and GitHub as examples — without those tasks leaving the machine for the cloud. When a task needs more reasoning than the local model can deliver, it falls back to frontier cloud models, making this a hybrid local/cloud system rather than a pure offline product.

Two capabilities are new relative to the Linux launch:

1. Local MCP connections. Portable Computer can connect to a user's own tools and app integrations through MCP (Model Context Protocol) running locally, rather than routing those tool calls through Perplexity's cloud. This is a concrete, shipping example of an MCP client embedded in a consumer product — the same host/client/server pattern explainx.ai's MCP guide covers, applied entirely on-device.
2. Scheduled tasks. Recurring work — Perplexity's own example is a daily stock report — can run on the PC while the user is away, without a person triggering each run manually.

Neither feature requires cloud round-trips once it's set up, which is the actual value proposition here: privacy and zero marginal token cost for local steps, not raw speed or a cheaper subscription.

Why this is genuinely useful, not just a demo

Local-first AI agents are a growing pattern, and Portable Computer on Windows is a real example a reader can act on this week if their hardware qualifies:

• No cloud dependency for local tasks means sensitive files and inbox content don't leave the device to get summarized or acted on — the same privacy thesis behind Perplexity's Mac hybrid compute work, applied here to a fully local Windows harness instead of selective routing.
• Local MCP support turns "connect my own tools" from a cloud-hosted integration into something that runs entirely on hardware the user controls — relevant to anyone already building a personal local AI system or running open-source models locally.
• Scheduled tasks running unattended is the kind of automation that usually implies a cloud server keeping something alive. Here it runs on a desktop that's already on, with no extra hosting cost.

Getting started is a single step: install or open the Perplexity Windows app on a PC with a qualifying RTX GPU. There's no separate download or beta gate mentioned beyond the app itself.

The honest cost critique

Community reaction on X raised a fair point worth stating plainly. One user, @JuakiGarcia, pushed back that users have to supply their own — expensive — GPU hardware and still pay a Perplexity subscription on top. That critique holds up:

• A qualifying 24GB RTX card is a $700-plus purchase for most people who don't already own one, on top of whatever Perplexity Pro, Max, or Enterprise tier is required.
• This is not "free local compute." It's a genuine hardware and subscription investment, and Perplexity's "unmetered" framing describes the marginal cost of local inference once you clear that bar — not the total cost of entry.
• The honest comparison is against a DIY local stack: explainx.ai's guide to closed-source AI vs. local open-source alternatives and the personal local AI system walkthrough both cover paths that skip the subscription entirely, at the cost of doing your own integration work instead of getting Perplexity's packaged harness.

Another user, @_mahns, asked whether mobile access or orchestrating these agents from anywhere is on the roadmap. Perplexity hasn't given a public answer as of this announcement — Portable Computer, on Windows as on Linux, is a single-machine, on-premises product. There's no indication yet of remote control or a companion mobile client that can trigger or monitor scheduled tasks running on the desktop box.

Who should actually try this

• Windows users who already own a 24GB+ RTX GPU — this is close to a free trial once you have a qualifying subscription; install the Perplexity Windows app and connect MCP integrations to what you already run.
• Anyone weighing a GPU purchase specifically for this — do the math against a used RTX 3090 (roughly $700-900) plus the subscription tier required, and compare that total against NVIDIA's DGX Spark path if you want more headroom for larger local models later.
• Laptop-only users — this isn't for you yet. Laptop RTX GPUs don't clear 24GB VRAM, so on-device inference isn't reachable without a desktop or dedicated tower.
• Teams wanting shared or remote access — not covered here. This is a one-operator, one-machine pattern, same as the Linux and DGX Spark releases.

Honest limitations

• 24GB VRAM is a hard floor, not a suggestion — most Windows PCs, including many marketed as "gaming PCs," do not clear it.
• Subscription is still required on top of hardware — this is additive cost, not a replacement for the subscription.
• No mobile or remote orchestration — confirmed unanswered as of this announcement.
• Cloud fallback means "fully local" is conditional — harder reasoning tasks still route to frontier cloud models, so privacy guarantees apply to local steps specifically, not every interaction.
• Feature parity with Linux/DGX Spark isn't fully detailed yet — Perplexity's announcement covers availability and the two new capabilities; a full changelog against the Linux release hasn't been published.

What builders should do this week

1. Check your GPU's VRAM before installing anything. Open Windows Task Manager's Performance tab, or NVIDIA's own control panel, and confirm dedicated GPU memory reads 24GB or higher before assuming the Perplexity Windows app will run on-device inference locally rather than routing everything to the cloud.
2. If you qualify, wire up local MCP first. Connect the tools you already use daily — Outlook, Gmail, GitHub — through local MCP before reaching for scheduled tasks, so you can verify the harness behaves correctly on your own files before trusting it with unattended, recurring work.
3. If you don't qualify, price out the upgrade against alternatives. A used RTX 3090 plus a subscription is one path; a fully DIY local stack built around open-weight models is another, with more setup work but no recurring subscription cost.
4. Treat "unmetered" as marketing shorthand, not a pricing claim. Read it as "no marginal token cost once you've cleared the hardware and subscription bar," not as "free."

Related on explainx.ai

• Perplexity Portable Computer: RTX GPU support confirmed on Linux
• Perplexity Portable Computer launches on DGX Spark
• What is MCP? Model Context Protocol complete guide
• Perplexity Mac hybrid compute: local PII routing
• Perplexity's Lily: a custom inference engine for Apple Silicon
• MacBook vs. dedicated GPU for local LLMs
• NVIDIA DGX Spark: best local LLM setup
• Build a personal local AI system
• Run open-source models locally in OpenCode
• Closed-source AI vs. local open-source alternatives

Official: Perplexity on X · Aravind Srinivas on X

Availability, hardware requirements, and feature details reflect Perplexity's September 14, 2026 Windows announcement. Verify current supported GPUs and subscription tiers against Perplexity's own documentation before buying hardware for this specifically.