How to Run an AI Agent Locally With Zero Token Costs

The zero-token promise is real, but narrow

Perplexity has released Portable Computer, a local version of its agentic Computer platform built with Nvidia. The idea is simple: the model, your files, and the agent run on hardware you already own, starting with Nvidia's DGX Spark desktop supercomputer and Linux machines with RTX GPUs. Work completed locally consumes no billing credits, and the system asks permission before sending any individual step to a more powerful cloud model.

The announcement is one of the most aggressive attempts yet to move serious agent workloads off the cloud. It is also easy to misunderstand. "Zero token cost" does not mean zero cost. It means you stop paying a vendor per step and start paying for hardware, electricity, setup time, and the opportunity cost of a smaller model.

What Portable Computer actually is

Portable Computer is the same Perplexity Computer interface moved to a fully local app. Perplexity's vice president of engineering described it as bringing the "exact same UI to a fully local app," including the agent harness and local inference. The bundling is the point: local models, the agent harness, an inference engine, tools, app connectors, and a security sandbox come as one system instead of forcing you to assemble vLLM, weights, and tools yourself.

The critical detail is the permission step. The agent does not silently phone home when it hits a hard task. It asks. That keeps sensitive data in one place and keeps the billing meter off unless you explicitly choose to turn it on.

The hardware you actually need

At launch the supported machines are Linux only. Windows support is planned for September. You need either an Nvidia DGX Spark or a Linux desktop with an Nvidia RTX GPU and at least 24 GB of VRAM — roughly a GeForce RTX 3090 or newer. That immediately rules out most laptops and every Mac. Perplexity called 24 GB the floor for delivering a good experience while still keeping the hardware broadly available.

Apple's same-day Mac Mini and Mac Studio announcement moves in a similar direction: local AI workloads on Apple Silicon. But Portable Computer is starting with Nvidia silicon, and Perplexity said it is "very focused right now on Nvidia hardware" when asked about Apple silicon. For a broader look at what local inference actually costs across platforms, see the guide on hardware for local AI models.

What stays local and what does not

The whole point of the product is that a task starts locally and stays there by default. That includes reading files, writing output, running local commands, and chaining steps together. Only when the local model is not enough does the system offer to escalate.

The launch demos showed both sides. In one, a retail investor had the agent review a folder of 1099s and investment documents locally — exactly the kind of sensitive financial material many users would hesitate to upload to a cloud service. In another, a founder analyzed a CSV of user funnel data locally and then pushed the finished analysis to Slack through Perplexity's connectors. Local-first does not have to mean disconnected.

Before any cloud escalation, the harness runs a PII classifier over what would leave the device and shows it to you. The remote advisor model returns text guidance only; it never touches your local files or tools.

The models you can run

At launch you can choose between Qwen 3.8 27B or PPLX 27B, a version Perplexity post-trained on its own harness. Nvidia's Nemotron 3.5 Lightning is listed as coming soon. The inference layer underneath uses vLLM, with an advanced mode for users who want to plug in their own endpoint.

The model selection matters because the harness is deliberately minimal to fit a 27B model: a succinct system prompt, a small set of core tools, and capabilities that load as on-demand "skills" rather than sitting permanently in context. Perplexity found that while Qwen 3.8 27B advertises a 260,000-token context window, it begins to struggle beyond 100,000 tokens, so the harness is built around that constraint.

Benchmarks: take them with the right salt

Perplexity published its own numbers on an internal Local Knowledge Work Bench of 53 tasks spanning deep research, financial analysis, and document creation. On that benchmark, Computer with Qwen 3.8 27B on a DGX Spark scored 82.6%, versus 77.6% for the open-source Pi harness and 74.0% for Hermes running the same model. Its post-trained PPLX 27B pushed the score to 85.4%.

The most interesting comparison is the hybrid one. On Terminal Bench 2.1, the fully local Qwen model scored 59.6% at essentially zero marginal cost. Letting it escalate to a Claude Opus 5 advisor in the cloud raised the score to 73.0% at about $0.415 per task. Running the frontier model alone scored 82.4% at $0.65 per task. The takeaway: escalation recovers roughly three-fifths of the gap to frontier performance at about two-thirds of the cost, and the user decides when that trade is worth making.

These are vendor-reported figures, so treat them as directional rather than final. But the shape is plausible: a good local agent plus selective cloud advice can be cheaper than running the frontier model for everything.

When a local agent is the right choice

Local inference makes sense in a few specific situations:

• Data sovereignty. The inputs never leave your machine unless you say so.
• Iterative work. Agents that loop dozens of times burn tokens fast. Local loops are "free" after the hardware is bought.
• Offline or air-gapped environments. You do not need an internet connection for the local steps.
• Fallback control. You know exactly when the cloud is invoked, and you can decline.

If none of those matter to you, a cloud agent is probably cheaper and easier. The hardware to run a capable local model costs more than a year of moderate cloud usage for most people.

The hidden costs no one quotes

The phrase "zero token costs" is technically true for local steps, but it is not the same as free. A DGX Spark is a serious piece of hardware. An RTX 3090 workstation is not a consumer purchase. Electricity, cooling, and the time to keep drivers, model weights, and the agent framework up to date are real ongoing costs.

There is also a capability tradeoff. A 27B local model is smaller than the frontier models Perplexity can call in the cloud. That means weaker reasoning on edge cases and a higher chance that the agent will need cloud fallback exactly when you wish it would not. For a deeper look at the limits of running models yourself, see how to run AI models locally.

How to set it up

You do not need a step-by-step tutorial to understand the shape of it. Portable Computer is available today for Perplexity Pro, Max, Enterprise Pro, and Enterprise Max subscribers on Linux. Install the app on a supported machine, sign in, and point it at a workspace. The agent can read files, run local commands, and produce deliverables in that workspace. When a step needs more power than the local model can provide, the app prompts you before sending anything out.

The important setup work is not the install. It is deciding:

• Which directories the agent is allowed to read and write.
• Which tools and connectors it can invoke locally.
• Your default policy for cloud fallback: always ask, ask above a cost threshold, or never.
• Whether the OS-level sandbox is required for your threat model. If it cannot run, the harness disables tool execution rather than running unprotected.
• How you will track whether the local model is good enough for each recurring task.

Start with one non-critical workflow and measure completion rate, speed, and how often it asks for cloud help. That single test tells you more than any announcement.

Sources and where to start

Perplexity announced Portable Computer on 25 August 2026 in a press briefing covered by VentureBeat. Nvidia concurrently signalled that it believes local AI has crossed from hobbyist curiosity to practical tool, which is why the chipmaker is co-promoting the launch as a reason to buy its local hardware.

• Perplexity partners with Nvidia to launch Portable Computer — VentureBeat
• Perplexity AI tool page on this site
• How to run AI models locally
• Hardware for local AI models

The honest summary

Portable Computer is a real change in where agent work can happen, but it is not a universal replacement for cloud agents. It is best for people and organizations that already have the hardware, care about data staying local, and run the kind of repetitive, looping agent tasks that make per-token pricing painful.

For everyone else, the launch is worth watching because it pushes the industry toward a hybrid model: do as much as you can locally, escalate to the cloud only when it is worth it, and keep the decision in your hands rather than the vendor's.