Perplexity Computer: Cloud, Local Mac and Windows Explained

Perplexity Computer is worth trying when a job spans research, files and a finished deliverable, especially if you already own hardware that can keep the AI work local. The October 20, 2026 Mac release makes the choice more useful: run in Perplexity's cloud, fully on a supported Mac, or locally on a supported AMD Windows PC. Choose by where your data needs to stay and how much work you can delegate, rather than by the promise of an autonomous agent.

What Is Perplexity Computer?

Perplexity Computer is an autonomous agent: software that plans and carries out multi-step work toward an outcome you specify. Give it a job such as preparing a client proposal, and it can gather context, work with files, create documents and use connected accounts. It is a service inside Perplexity, not a physical computer you need to buy. Perplexity's Computer documentation describes that distinction.

Think of it as delegating to an assistant with a workroom. You supply the assignment, relevant materials and authority to act. The workroom might be in Perplexity's cloud or on your machine. That location changes which information travels and which computing resources do the work.

How Does Perplexity Computer Work?

Computer turns your requested outcome into steps, chooses tools and can delegate parts to specialist agents. In its normal cloud mode, it executes inside an isolated cloud environment. The useful difference from a search answer is the chain: gather material, analyze it, create an output, then perform an authorized follow-up. Perplexity explains the execution model here.

Your instruction still needs a finish line. “Prepare a proposal and leave it for review” is decidable. “Grow my business” leaves the agent guessing about priorities, acceptable commitments and when to stop.

Three Ways to Run It, With Different Data Boundaries

The biggest change is that a desktop app can now run the AI itself locally, as well as access local files. Perplexity announced fully local Mac execution on October 20, 2026, following AMD Windows support on September 24.

The following comparison was verified against Perplexity's live pages on October 21, 2026. These are the three routes relevant here; supported NVIDIA and Linux options also exist.

The distinction comes from the Mac execution page, AMD announcement and current credit rules. “Local files” alone tells you where the inputs are stored. It does not tell you where the model processes them.

Cloud is the easiest starting point when the job depends on public research or information you are comfortable processing remotely. Open Perplexity, select Computer, then Start a task. The web route needs no local model installation. A founder assembling a public competitive brief can start there without buying a new machine.

Perplexity Computer Mac: Fully Local or Hybrid

Fully local Mac execution is available to Pro, Max and Enterprise subscribers on supported Apple silicon Macs. Its recommendation is 36 GB or more for fully local Computer. That is a recommendation, not a published hard minimum. Mac launch details.

Hybrid Compute starts in the cloud. Cloud models handle research and reasoning while a model on your Mac processes private files. Perplexity's help center documents 24 GB or more of unified memory for that route; the Mac app requires macOS 15 Sequoia or later. A 24 GB Mac therefore has a documented Hybrid path. Do not turn that fact into a promise about fully local performance. Hybrid requirements.

For an operator with confidential notes and an internal rate sheet, fully local is the relevant option. If the same job needs current market research, Hybrid is a different, deliberate choice about how to divide the work.

AMD Windows: Check the Processor and Memory Type

Portable Computer brings the model and agent onto supported AMD Windows PCs. Perplexity specifies an AMD Ryzen AI Max Series processor, including the Ryzen AI Halo platform, at least 24 GB of GPU-accessible memory, Windows 10 or 11, and approximately 20 GB of free disk space for initial setup. Pro and Max subscribers on individual and enterprise plans are eligible. Official AMD requirements.

An ordinary Windows laptop with 24 GB of RAM does not qualify on that number alone. Install the Perplexity app, download a model through Settings > Local Inference, then select it under Local in the composer. Our Portable Computer AMD Windows guide covers the setup in detail.

For recurring work, the app must remain open and the PC awake. That makes a regularly closed travel laptop a poor home for a job another person expects every morning.

Plan access: Pro is the consumer entry point; Perplexity's current plan page displays $17/month when billed annually for Pro and $167/month when billed annually for Max, with Enterprise Pro and Enterprise Max also eligible for Computer.

Perplexity Computer Credits: What a Run Consumes

Perplexity says work handled by a local model consumes no Computer credits. Cloud execution has a usage meter, and the size of the finished file is no guide to the amount of work behind it. Perplexity gives 350 to 950 credits as an example range for a multi-step research report and 2,400 to 9,800 for a data dashboard. Those are vendor examples, not quotes for your task. Official credit examples.

There is also a current Pro distinction: paid Consumer Pro accounts are gradually receiving weekly included Computer usage. Check for Included in Pro on your Usage page. It is separate from credits and is consumed first. If you have included usage but no credits, tasks use the Light effort setting. Once included usage runs out, Computer switches to available credits or your enabled auto-refill; otherwise the session is blocked until you add credits or the allowance resets. Auto-refill is off by default. Included-usage rules.

Before repeating a workflow, inspect its thread's Usage dropdown. The full allocations, purchase rules and conversion belong in our Perplexity pricing and credits guide. A subscription grants access; it does not make every cloud task unlimited.

What It Can Do With Your Files and Accounts

A folder permission grants continuing working access inside that folder. Perplexity says the app can find, read, create and change those files without requesting access each time. Removing the folder from a request only deselects it; on Mac, revoke access through Settings > Permissions. Local-file permissions.

For a founder preparing a proposal, that suggests a focused project folder containing copies of the relevant inputs. Giving access to the whole company drive creates more room for irrelevant context and unintended changes without improving the assignment.

Connected accounts add another kind of authority. A connector is an authorized connection to an app such as email or a document service. Computer can use that connection to retrieve information and, where supported, act on it. Manage these through Plugins, or Customize > Connectors where that interface still appears. Connection controls.

The most useful caveat is stated directly on Perplexity's Mac page: fully local describes where the AI runs; sending a message or meeting invitation still uses that app's service. A proposal can be prepared locally and then leave through email. Those are separate steps with separate consequences. Local processing and app services.

In Hybrid mode, Privacy Gate can keep content local, mask personal details before cloud use or request sharing approval. That gives you controls over a transfer. It does not make every hybrid task an entirely local task.

How to Use Perplexity Computer for a Client Proposal

A client proposal is a good fit because the work has bounded inputs and a reviewable ending. Perplexity itself describes using confidential discovery notes, a client brief and an internal rate sheet to draft scope, build an Excel budget and prepare an Outlook cover email. The procedure below expands that documented Mac example; it is an illustrative workflow, not a report of a test performed here.

1. 1

   Prepare the inputs and choose the execution mode

   Create a project folder with the discovery notes, approved client brief and current rate sheet. Use copies so the originals remain available for comparison. On a supported Mac, open the latest Perplexity app, complete the in-app local-model setup and choose fully local before starting. Grant access to that project folder.

2. 2

   Specify the deliverables and stopping point

   Ask for a scope draft, a budget workbook and an email draft. Include the rules that would otherwise live in your head: use only approved rates, flag missing quantities, distinguish client requests from your recommendations, and leave commitments for review.

3. 3

   Let it prepare the package

   The intended sequence is to extract requirements, map them to deliverables, calculate the budget from the supplied rates, then prepare a cover note. Ask it to name the source file behind each requirement and list unresolved assumptions. A polished document without those references is harder to check.

4. 4

   Review before anything goes out

   Compare scope against the discovery notes, inspect workbook formulas and quantities, and check the recipient and promises in the cover note. Treat an invented deadline, an unsupported discount or a missing deliverable as a failed draft that needs correction. Save the approved package separately.

5. 5

   Decide separately about research and delivery

   If a claim needs current web evidence, decide whether to authorize a cloud step or supply the evidence yourself. If the draft is ready, send it through your normal approval process. Preparing the package locally does not make the later email delivery local.

A useful starting instruction is:

Work only with the files in this project folder. Draft a client proposal and an Excel budget using the supplied brief, discovery notes and rate sheet. Save the drafts here and write the cover email as a local text file. Reference the input behind each deliverable and mark missing information as an open question. Do not change the source files or send anything. Keep the task fully local; if it needs a cloud step, stop and explain why.

This version saves the email as text until review, making the first run easier to inspect without connecting an inbox. The acceptance test is the package: correct scope, traceable inputs, correct calculations and an explicit list of open questions. There is no task-specific credit estimate here because this workflow was not executed.

Should It Replace the Agent You Already Pay For?

Keep the existing agent unless Computer removes a specific limitation in your workflow. The strongest reason to try it is a combination you need: multi-step research and deliverables, or local AI execution on hardware you already own. A new interface alone does not justify moving working processes.

Perplexity Computer vs Claude Cowork

Claude Cowork, Anthropic's agent for knowledge work, already handles tasks across files and tools. Claude Code is its coding agent for repository work such as building, refactoring and testing. Those are the distinctions in Anthropic's product guide.

If Cowork already produces the proposals you need, use it as the baseline. Give Computer an equivalent permitted input set and compare the corrections you must make, the external access required and the recorded usage. Switch only when the result improves a constraint that matters to you. No benchmark in the documentation establishes that Computer will write your proposals better.

For a technical builder whose paid agent already edits and tests a repository effectively, keep that setup unless broader business tasks are the missing piece. Our Claude Cowork vs Claude Code comparison explains that workload boundary; the Perplexity review covers the broader research product.

Perplexity Computer Use Cases

Operators should try it on a recurring preparation task: a proposal pack, a review queue or a status brief whose inputs and output can be named. The consequence is less manual transfer between documents, provided the review burden stays manageable.

Founders should start with an existing subscription and machine. A public research brief can use cloud Computer; a private proposal assembled from local materials is a reason to investigate local mode. Buying hardware before establishing the recurring job reverses the decision.

IT buyers should evaluate the execution and permission model first. Perplexity provides administrator controls for local inference and, on Mac, rules governing local versus cloud handling. Those controls matter more to a rollout than a broad promise of autonomy. Managed-account controls.

What's Overhyped

“Local” does not mean a free subscription, unlimited capability or no network activity. It describes where a particular part of the work runs. External app actions still use services, and an approved cloud step changes both the data path and usage accounting.

“Desktop agent” also does not imply identical app control everywhere. Perplexity's help center says installed-app control is not yet available on Windows. Portable can use local tool adapters called MCP servers, but its docs explicitly say those servers run outside the sandbox with your user permissions. Platform differences, local-tool boundary.

Finally, a completed task is not an approved business decision. The client can still receive an inaccurate scope or an unprofitable commitment if nobody reviews it. The practical value is a smaller, clearer review job, not permission to stop owning the outcome.

Frequently Asked Questions

What is a Perplexity Computer for?

It carries a multi-step task from source material to an output, such as researching a market, preparing a document or assembling a client proposal. It is most useful when a plain answer would still leave you doing the work across files and apps.

Is a Perplexity Computer worth it?

Try it when it can remove a specific workflow limitation, especially if you already have an eligible plan and supported local hardware. Keep your current agent when it produces acceptable work with less review and setup.

How much does a Perplexity personal computer cost?

Personal Computer is software accessed through an eligible paid Perplexity plan, with hardware and chargeable cloud usage separate. The annual-billing rates are given above; our Perplexity pricing guide covers the full plan and credit details.

Is a Perplexity Computer an actual Computer?

No. Computer is Perplexity's agent software. Local modes use supported hardware you own; the cloud mode runs remotely.

Can a Perplexity Computer run on Windows?

Yes. The desktop app supports Windows, but local AI execution needs qualifying hardware. The AMD route requires Ryzen AI Max Series hardware and the GPU-accessible memory specified above; installing the app alone does not enable it.

Is Perplexity Computer a security risk?

Its access is consequential: it can read or change permitted files and act through connected services. Scope the permissions to the assignment, review outbound actions and check whether a requested cloud step fits your data rules.

What are the downsides of Perplexity?

For Computer, the main tradeoffs are variable cloud usage, hardware requirements for local inference, setup for connected tools and the need to review results. Those can outweigh the benefit for a simple question or an already effective workflow.

Is there a difference between Perplexity and Perplexity Computer?

Perplexity Ask finds and synthesizes answers. Computer carries out tasks with tools and produces deliverables. Ask searches do not consume Computer credits.

Does Perplexity Computer control your Computer?

Through the desktop app, it can work with permitted files and supported apps. The scope differs by platform and connection. Granting local file access is separate from choosing a model that runs locally.

Your Monday Move

Choose one proposal you already need to prepare. Put copies of its approved inputs in a dedicated folder, select the execution mode deliberately, and ask for drafts only. Record the corrections and any cloud approvals before repeating the workflow. If your current agent needs less intervention, keep it. If local Computer removes a meaningful data-processing constraint and produces an acceptable package, expand from that evidence.

For practical notes on choosing and using AI at work, join the newsletter.