Perplexity's Portable Computer Runs AI Agents Locally With Zero Token Costs

Introduction: The First Local-First Agent From a Major Lab

Most AI agents in 2026 have the same fine-print problem: every step they take is metered. Every file read, every planning step, every tool call burns tokens on someone else's server — and your documents go along for the ride. On August 25, 2026, Perplexity launched Portable Computer in partnership with NVIDIA, and it inverts that model completely. It's a fully local version of the company's agentic Perplexity Computer platform: the model, your files, and the work itself can all stay on your machine, and work completed locally consumes zero Perplexity credits with essentially zero marginal token cost.

This isn't a hobbyist project running a quantized model in a terminal window. It's a major AI company shipping an entire agentic control plane — orchestrator, planner, tool router, scheduler, durable task queue, and local search index — that runs on-device by default and only touches the cloud when you explicitly allow it. Running a model locally has been possible for two years. Running the whole agent locally is the new part.

What Portable Computer Actually Is

Portable Computer is the local-first sibling of Perplexity Computer, the company's agentic platform that can research, plan, and execute multi-step workflows. The key difference is where execution happens. Tasks start locally by default: a local model reads your files, plans the work, and runs the tools on your hardware. The agent doesn't ping a frontier cloud model unless a step genuinely needs one — and it has to ask your permission first.

At launch, users can run two models locally: Qwen 3.8 27B, the open-weight workhorse, and PPLX 27B, Perplexity's own post-trained variant tuned for agentic work. NVIDIA's Nemotron 3.5 Lightning, a speed-optimized model, is slated to join the roster next. A 27-billion-parameter model won't out-reason a frontier flagship, but for the bulk of everyday agent work — summarizing documents, drafting emails, restructuring files, searching your own data — it's squarely in the "good enough, and it never leaves this machine" zone.

The Full Local Stack: More Than a Model

The genuinely novel part of the announcement isn't the model — it's everything around it. Portable Computer bundles a complete agentic stack that executes on-device:

• Model inference — Qwen 3.8 27B or PPLX 27B running on your GPU.
• Agent harness — the orchestrator, planner, and tool router that turn a request into a multi-step plan.
• Security sandbox — tool execution is contained, so an agent that misbehaves can't wander your filesystem.
• Local file access — the agent reads and edits your documents directly, with no upload step.
• Local search index — your own corpus, searchable without a round trip to anyone's servers.
• Connectors — built-in integrations for Gmail, Google Drive, Slack, and GitHub, linking the local agent to the services your work already lives in.

That last combination is what makes this practical. A term sheet can sit in your local files while the agent drafts comparison tables from it — and if you decide you want current market comps from the live web, the orchestrator can fetch them, with your approval, without ever shipping the confidential document itself to the cloud.

The model running locally was never the hard part. The orchestrator, sandbox, and connectors deciding things locally — that's what changed this week.

How Cloud Escalation Works (and Its Guardrails)

"Zero token cost" comes with an asterisk, and Perplexity has been unusually specific about the guardrails. Local-only is the default mode, and moving a task to the cloud requires an explicit, per-action approval from you. Content inside a local document cannot independently trigger an escalation, and it cannot override product controls. You have to manually enable an "allow advisor escalation" setting in the app's preferences; without it, no computational work proceeds to the cloud at all.

There are two more design details worth noting. Escalation is restricted to a single instance per request, preventing automatic propagation across the rest of a task or into future sessions — so approving one cloud lookup doesn't silently authorize a chain of them. And the model itself, through post-training, determines whether a step genuinely requires a frontier model, which keeps the default path local instead of nagging you toward the cloud at every step. It's a privacy architecture that assumes the user is the only trusted party in the loop.

The Hardware Reality Check

Now for the catch, and it's a big one: Portable Computer launches on Linux, requiring NVIDIA hardware with at least 24GB of VRAM — the NVIDIA DGX Spark and compatible RTX systems running DGX OS or Ubuntu, roughly an RTX 3090 or newer. Windows support is planned for September 2026, and there's no macOS version yet. It's available to Perplexity Pro, Max, Enterprise Pro, and Enterprise Max subscribers.

So no, this isn't something most people install on a $700 laptop today. The entry ticket is a serious GPU or a DGX Spark box. Analysts have been quick to point out that the hardware barrier to entry is substantial — you're trading a per-token subscription burn for a significant upfront capital cost. But for people who use AI agents all day, the math can work: heavy users routinely burn tens of dollars a day in API and credit costs, which compounds against a one-time GPU purchase quickly. And enterprises that couldn't put sensitive documents in front of a cloud model at all suddenly have an agentic option that compliance can live with.

What Local-First Agents Mean for the AI Tools You Pick

Portable Computer is part of a wider local-first turn in 2026 — open-weight releases like Meta's Muse Glimmer 30B and Qwen3.8-27B have made capable local models commodity, and now the agent layer is following. If you're evaluating AI tools, here's how to read the trend:

• Privacy becomes a feature you can buy, not a policy you trust. Tools that keep execution on-device can make guarantees cloud tools can only promise. For legal, healthcare, and finance work, that's a category unlock.
• Watch pricing models split. Zero-marginal-cost local execution puts pressure on per-seat and per-token pricing. Heavy users should start calculating whether a local agent plus selective cloud escalation beats an all-cloud subscription.
• The hybrid pattern is the real winner. Local for the sensitive 90% of work, frontier cloud models for the hard 10% — that's the architecture Portable Computer formalizes, and it's the one to look for in other tools.
• Check the escalation controls. Any tool that mixes local and cloud execution should tell you exactly what triggers a cloud call and what data crosses the boundary. Per-action, user-initiated approval is the bar to demand.

Compare the cloud and local players side by side in our AI Agents and AI Productivity Tools categories to see which assistants can already run where your data lives.

The Bottom Line

Perplexity's Portable Computer won't change what most people do this week — Linux, a 24GB GPU, and a paid subscription is a narrow doorway. But it's a meaningful inflection: the first time a major AI company has shipped a complete, sandboxed, connector-equipped agent that treats your own hardware as the primary runtime and the cloud as an opt-in accelerator. Zero marginal token cost and default-local privacy attack the two biggest complaints about agentic AI — the meter and the trust gap. If the Windows release lands in September and the model roster grows, local-first agents stop being an enthusiast niche and start being a procurement question. The tools that offer a credible local path will increasingly win the work that matters most.

Frequently Asked Questions