How to Use Perplexity Portable Computer

You can now put a recurring file-analysis job on a supported AMD Windows PC and let Perplexity run the model and agent locally, without spending Computer credits on the local steps. The useful question is not whether Portable Computer sounds private. It is whether your machine qualifies, the task stays inside one permitted folder, and you know exactly when a cloud approval changes the cost and data boundary.

The Short Answer

Use Perplexity Portable Computer this way:

1. Confirm that you have Pro or Max, Windows 10 or 11, a Ryzen AI Max Series system with at least 24 GB of GPU-accessible memory, and roughly 20 GB of free storage.
2. Install and sign in to the Perplexity app for Windows.
3. Open Settings > Local Inference and download PPLX 27B or Qwen 27B.
4. Open Settings > Permissions, enable local filesystem access and Desktop tools, then permit one test folder.
5. In the composer, choose the downloaded model under Local.
6. Run a small file task with known answers before you schedule anything important.

That is the complete path. A Windows PC with 24 GB of ordinary system RAM does not automatically qualify. The September 24 AMD requirements specify a Ryzen AI Max Series processor and at least 24 GB of GPU-accessible memory. Perplexity's AMD announcement is the current source for that route.

Portable Computer Is the Local Brain, Not Just File Access

Portable Computer and Personal Computer solve different halves of the job. Think of Personal Computer as the key to a filing cabinet: it gives Computer tools to find, read, create, and edit local files. Portable Computer moves the analyst into the room: the agent runtime and model inference also run on your device.

You can use local file tools without turning on Portable Computer. You can also use both together, which is the setup this guide uses. Perplexity says the local model, planner, tool router, scheduler, durable task queue, file search index, and sandbox run on the PC. The help page explains the distinction directly.

That boundary matters because a task can still ask for the cloud. Perplexity Search, wide research, deep research, browser use, connected apps, or harder reasoning through more than 15 frontier models can be brought into a local task. Portable Computer asks before it sends information from the device. If you approve, that step runs in the cloud and uses Computer credits at the standard rate.

Check the Machine Before You Download Anything

For the new AMD path, all five rows below need to be true.

There is a documentation mismatch worth seeing before it wastes an afternoon. The main help article says it was last modified September 15 and still lists DGX Spark or supported NVIDIA RTX hardware. Perplexity's dated September 24 post adds Ryzen AI Max support. Use the newer announcement for AMD eligibility, while using the help article for the Windows setup sequence.

The app and local inference are not free merely because the local task consumes no credits. Perplexity's consumer pricing page currently displays Pro at $17 per month and Max at $167 per month when billed annually. Hardware, electricity, review time, and any approved cloud work remain costs.

The useful budget equation is:

Monthly local workflow cost = paid plan + hardware amortization + review time + approved cloud credits

If you already own a qualifying PC and already pay for Pro or Max, the extra Computer-credit cost of a fully local run is zero. If you need new hardware, do not force a payback claim from an unspecified purchase price. Measure the job first. As an outside reference, InvoiceReconcile lists paid reconciliation plans from $19 to $99 per month. That is not the same product, but it is a real alternative budget line for this particular job.

Set It Up Without Widening the Permission Boundary

The safest first setup is one model, one folder, and one task with answers you already know.

1. Install the Windows app and sign in

Install the Perplexity app for Windows. Portable Computer is not a second Windows installer. Local inference is a setting inside the app.

2. Download one supported local model

Open Menu > Settings > Local Inference. Under on-device models, download PPLX 27B or Qwen 27B. The inference engine and sensitive-content classifier install with the model. Only one local model runs at a time, so choose one for the first test instead of treating two downloads as a prerequisite.

3. Create one fixture folder

Create a folder called Portable-Reconcile-Test. Put only synthetic test files in it. Do not start with a live inbox, payroll export, contract repository, or shared drive.

4. Permit only that folder

Open Settings > Permissions. Enable Local file system access and Desktop tools, then add only Portable-Reconcile-Test. The sandbox controls file access to permitted folders. A smaller permission boundary also makes a bad instruction easier to detect and reverse.

5. Select the local model before the task

Return to the composer, open the model menu, and select the downloaded model under Local. Seeing a local model selected is the provider's check that the task will start on-device. A cloud model selected in the composer is a different execution path.

6. Decline cloud escalation during the fixture run

The test below does not need current web data, a connector, or frontier reasoning. If an approval prompt appears, decline it and record why the task asked. A clean local test should not require a cloud step.

Run a Reconciliation With Three Known Exceptions

This fixture is designed to fail in exactly three known ways. It gives you something better than a vague demo: a result you can mark right or wrong.

Create contract-rates.csv with these four rows:

Create invoices.csv with these five rows:

Then give the local model this instruction:

Work only inside the permitted Portable-Reconcile-Test folder. Compare invoices.csv with contract-rates.csv. Write exceptions.csv with exactly one row per invoice ID that has a duplicate, no matching contract rate, or an amount variance. Use the columns invoice_id, issue, expected_amount, invoice_amount, and difference. Do not use the web, connected apps, or a cloud model. If any step asks to leave this folder or use the cloud, stop and explain why.

The expected exception file has three findings:

Check the row count, each invoice ID, and each amount. Do not score the run by how polished the explanation sounds. The output is correct only if it catches those three exceptions without adding a fourth.

This procedure is reproducible, but it was not executed in this article. No supported Windows machine and authorized paid account were available for a hands-on run, so there is no invented setup time, completion time, accuracy percentage, credit delta, or network measurement here.

Schedule the Same Task, Then Keep the Machine Available

Schedule the reconciliation only after the one-off run passes. Reuse the same folder, same instruction, and same expected schema. Perplexity says scheduled work uses the local scheduler and durable task queue, but the app must stay open and the PC must stay awake.

That last condition is operational, not cosmetic. A sleeping laptop is not a dependable morning control. If the exception file matters to a close process, use a machine with an explicit power policy, name an owner for failed runs, and keep the manual check until the schedule has proved reliable.

Record these fields for the first real test:

An unchanged credit balance would support the narrower claim that the run did not consume Computer credits. It would not prove zero network activity. If you need that assurance, use network monitoring and an approved security test rather than treating the product label as evidence.

Six Workflows That Benefit Most

The best use cases are repetitive, file-heavy jobs where a human can review a compact exception output.

The common shape is exception-first automation. Portable Computer is less attractive when every task needs live web data, the machine cannot stay awake, or nobody owns the review step.

Three Things Worth Building

1. A local invoice-exception runner

This is the strongest opportunity. Build a small deployment package for controllers and bookkeeping firms that installs the folder structure, validates CSV columns, runs the reconciliation prompt, writes a stable exception schema, and keeps a run log.

The demand is specific and commercial. US Google data shows about 170 monthly searches for invoice reconciliation software, 0 keyword difficulty, $51.67 CPC, and 24% yearly growth. Existing specialist plans run from $19 to $99 per month, which gives buyers a real comparison point.

The smallest sellable version needs three things: a fixture-based acceptance test, reusable mappings for two CSV exports, and a human review screen for exceptions. The catch is accounting variability. Column names, currencies, partial payments, and approval rules differ by customer, so a generic prompt is not the product. The mappings and evidence trail are.

2. A private contract-clause checker

Package a permitted-folder workflow for legal operations teams: one clause checklist in, a contract set in, and a cited exceptions table out. The buyer pays for faster issue spotting, not autonomous legal judgment.

US Google data shows about 590 monthly searches for ai contract review software, 27 keyword difficulty, $60.75 CPC, and 2,018% yearly growth. That is more demand than the invoice niche, but the trust and liability burden is much higher.

An MVP can support one contract type, one approved checklist, clause-level source references, and mandatory lawyer review. The catch is decisive: a missed clause can cost far more than the software. Without strong evaluation sets and a visible review trail, this should not ship as a decision maker.

3. An AMD readiness and runbook kit

Build a paid setup service or lightweight verifier for IT teams adopting local agents on Ryzen AI Max PCs. It would check the supported processor, GPU-accessible memory, free storage, app version, local model state, folder permissions, power policy, and a fixture result.

The nearest demand signal is smaller but rising. how to build a local ai agent shows about 30 US searches per month, 9 keyword difficulty, $12.70 CPC, and 400% yearly growth. The eight organic results for this article's target query include no independent AMD walkthrough from eligibility to a scheduled file task.

The MVP is a readiness report plus the exact three-exception test in this guide. The catch is platform risk. Perplexity can improve its own setup diagnostics, and a 30-search market is not a broad software category. This works better as a wedge into local-agent implementation than as a standalone SaaS bet.

Where Portable Computer Stops Being the Right Choice

Do not use it just because “local” sounds safer. The hardware route is narrow, the plan is paid, the model download is substantial, and scheduled work depends on an awake PC with the app open. Only one local model runs at a time. Brain and saved memory are not available in local mode, and Perplexity describes the local model as smaller than a frontier model.

The security boundary also needs precise language. Folder access is controlled and code execution is sandboxed, but a local MCP server runs outside that sandbox with your user permissions. Connected apps, browser use, web research, and approved cloud reasoning can cross the device boundary. Portable Computer asks first, which is a useful control, not proof that the machine makes no network calls.

Choose another route when you need an air-gapped guarantee, an always-on shared server, a multi-user policy layer, unsupported hardware, or a workflow that constantly depends on the web. For the wider product decision, see the Portable Computer alternatives comparison. Its August 26 hardware section still frames the product as DGX Spark-only with Windows still ahead, so that section needs a freshness update and should not be used for current AMD eligibility. The broader Perplexity review is the better branch when your question is whether Perplexity fits your research work at all.

The Monday Move

If you already have a qualifying Ryzen AI Max Windows PC and Pro or Max, reserve a focused session next week for one bounded pilot. Permit only Portable-Reconcile-Test, download one model, run the three-exception fixture, and record the result. Schedule it only after the output is exact. If the machine fails the eligibility check, stop there. Do not buy hardware until a real recurring workload and its current labor cost justify the budget.

Can I run a Perplexity Computer locally?

Yes, through Portable Computer on supported hardware. The new AMD Windows route requires Windows 10 or 11, a Ryzen AI Max Series system with at least 24 GB of GPU-accessible memory, a Pro or Max plan, and approximately 20 GB of free storage for initial setup.

What is a Perplexity Portable Computer?

It is Perplexity Computer with the model inference and agent stack running on your own supported machine. It is software inside the Perplexity app, not a separate physical computer.

How do I access my Perplexity Computer?

Install and sign in to the Windows app, open Settings > Local Inference, download a supported model, set local file permissions, then choose that model under Local in the composer.

How much does Perplexity Computer cost per month?

Portable Computer requires Pro or Max. Perplexity's consumer page currently displays Pro at $17 per month and Max at $167 per month when billed annually. The qualifying hardware, electricity, operator review, and any approved cloud credits are separate costs.

Is it safe to use a Perplexity Computer?

It provides useful controls: a local sandbox, permitted folders, a sensitive-content classifier, and an approval prompt before cloud transfer. Safety still depends on narrow permissions, trusted MCP servers, careful review, and whether your policy allows the approved cloud and connector paths.

If you want a bounded local workflow like this built and tested for your business, I can help with the automation design.