Perplexity Computer vs Claude Cowork: What Actually Stays on Your Mac?

An AI agent can open a spreadsheet on your Mac while doing its thinking somewhere else. If that spreadsheet holds client finances, private correspondence or a confidential proposal, the distinction is worth understanding before you connect the folder.

On October 20, 2026, Perplexity introduced fully local Computer for Mac. It says the model, agent coordinator and task execution can run on supported Apple silicon hardware. By contrast, new Claude Cowork tasks on Pro and Max run on Anthropic's servers, even when they work with permitted files on your Mac. Cowork is being folded into the broader Claude experience, so the separate selector may no longer appear.

The short answer: If you need to keep a document away from cloud AI inference, Perplexity's fully local mode is the relevant route of these two, provided your Mac supports it and you keep the task local. If cloud processing is acceptable, both Perplexity Computer and Claude offer agents that continue working after your laptop closes. Then compare the models, connectors, workflow and usage allowances you actually need, rather than declaring a winner based on where the cloud agent runs.

This is an evidence-based comparison of published capabilities, not a Choosely hands-on benchmark or network-traffic audit. For the broader platforms, see Choosely's Perplexity and Claude catalog profiles.

Four routes to the same Mac

Perplexity's product family spans cloud Computer, Hybrid Compute and the new fully local mode. Claude's current Pro/Max Cowork route uses cloud execution, with desktop access when a task needs local resources. Each route can touch Mac data under different conditions.

Perplexity's September Hybrid Compute announcement documents the 24 GB minimum, while its Hybrid product page specifies 32 GB for best results. The October 20 launch separately recommends 36 GB or more for fully local work. Anthropic documents its cloud and local Cowork architectures. These memory figures describe different modes; they should not be blended into a single requirement.

Perplexity's three choices

Personal Computer brought Perplexity into Mac folders and native apps before the latest launch. That did not automatically put the AI model on the device. Hybrid Compute then split work between cloud reasoning and local handling of selected private data, with a privacy gate that can redact sensitive material, keep a step local, require approval or refuse it.

The October release puts the agent coordinator and AI inference on the Mac for fully local tasks. The launch names Qwen 3.6 and PPLX Pro as local choices. It does not consistently specify the underlying Qwen version used to build PPLX Pro across its communications, so that lineage is not needed to make the buying decision. Perplexity offers examples of reviewing local financial data, building a proposal from confidential notes and updating spreadsheets or presentations.

There is a clear exit from the local boundary: Perplexity says users can authorize cloud escalation for current information, browser use, connected apps and harder reasoning through one of more than 15 cloud frontier models. The content needed by an approved cloud step is then handled outside the Mac. Likewise, sending an email, calendar invite or message uses that external service even if drafting happened locally. An agent's permission choices matter task by task.

Claude can work locally without running a Claude model locally

Anthropic announced the unified Claude experience in September. For new Pro and Max Cowork tasks from October 6, the agent loop and code execution run on its servers. Anthropic removed the “Only on your computer” option on those plans and says users cannot switch new tasks back to that mode.

There are real exceptions. Existing tasks started locally before October 6, and older local scheduled tasks, can stay in their original mode. On Team and Enterprise, administrators can allow local desktop Cowork while disabling cloud sessions. But locally executed Cowork is not a locally hosted Claude language model: it means the agent loop, file operations and isolated code environment run on the device, while model inference still requires a remote Claude service. Claude Code's local data-flow documentation confirms that local execution sends prompts and model outputs over the network to the model provider.

For a new Claude cloud task, a folder you authorize stays on your Mac until specific files are requested. Claude accesses the permitted file through the open desktop app and processes a copy on Anthropic's servers; it can also write changes back into connected folders. According to Anthropic's current access rules, that local-folder bridge works while the desktop app is open for a session initiated on desktop or a scheduled task. A fresh cloud session started directly from the web or phone cannot simply reach the Mac's folder under that rule, even if you later open the app. Anthropic says copies fetched by a session are deleted under its retention process when the session is deleted.

The cloud task itself can keep processing after the laptop closes, but it cannot fetch another Mac-only file while the desktop is offline. Team and Enterprise controls and legacy local deployments need separate treatment from a consumer Pro or Max subscription.

What actually leaves the Mac?

The table below separates the stages most easily confused. “Unknown” means the vendor documentation reviewed on October 22 does not establish an end-to-end answer for that mode; it does not imply the activity never occurs.

Sources include Perplexity's Mac launch, Personal Computer setup and Hybrid description, alongside Anthropic's architecture and computer-use safety guidance.

A sandbox and an access permission are different controls. Claude says its cloud code runs in an isolated sandbox, while desktop computer use can click in approved apps and has no sandbox between the agent and what appears on screen. It takes screenshots to understand the interface. Perplexity likewise describes sandboxed tool execution, but its Mac setup requests substantial OS privileges. Either product can make an unwanted change within its approved authority.

Check your Mac and its permissions before subscribing

Start with Apple menu → About This Mac. Apple's Mac support documentation says this displays the chip, macOS version and installed memory. The memory numbers matter:

• Under 24 GB: Perplexity's documented Hybrid minimum is not met. Fully local suitability also has not been established. Cloud options do not demand on-device model inference.
• 24–31 GB: Hybrid meets the published memory minimum on compatible Apple silicon hardware and macOS 15+, although Perplexity recommends 32 GB for the best hybrid experience.
• 32–35 GB: Hybrid meets its preferred memory tier. Fully local's 36 GB recommendation has not been met, but that figure is not an announced hard cutoff.
• 36 GB or more: Meets Perplexity's stated recommendation for fully local, subject to supported hardware, app entitlement and model setup. Performance is still untested by Choosely.

The Personal Computer installation guide requires macOS 15 or later for that Mac app and asks for Accessibility, Screen & System Audio Recording and Full Disk Access. Accessibility enables clicking and typing; screen access lets the agent see the display; Full Disk Access is a broad permission to read files. Users then choose folders and sign into services. The guide routes web tasks through Comet. The October 20 fully local post does not restate an independent OS requirement or publish exact model download sizes, free-disk requirements or a separate hard RAM minimum.

Those permissions deserve attention if the machine contains records beyond the assigned task. Connect limited working folders, use copies of sensitive originals and review what an app can see before accepting prompts for file changes or external actions.

Does fully local mean offline?

Perplexity says a supported task's model and agent can run on the device. Some launch reporting describes use without Wi-Fi. The company has not supplied a complete Mac-local offline support specification covering initial authentication, downloaded models, app startup, updates, tools and every task type. A specific downloaded, local-only workflow may run offline, but that is not a blanket documented guarantee for everything Computer can do. Choosely has not disconnected a Mac to test it.

Price and usage: $20 access is not the same as unlimited work

US web subscription reference prices, checked October 22, 2026:

These are published US web prices and allowances, not app-store billing or a personal entitlement check. Use Perplexity pricing, its Computer credit rules and its description of Pro's weekly Light allowance; compare Claude pricing. Perplexity's bonus credits are granted once per account and do not reset with a new monthly bill. Max's 10,000 plan credits do refresh monthly. Check your actual account before assuming a promotional bonus is available.

Fully local Mac billing has one important update. In Perplexity's official company announcement about the October 20 launch, the company says PPLX Pro local tasks do not consume Computer credits. This is an official company claim, stronger than an inference from its older NVIDIA/AMD local products. However, the credit help page does not yet spell out that Mac exception, Choosely has not measured a session's debit, and Perplexity has not clarified in those materials whether the same rule applies to every alternative local model or to Pro's weekly included allowance.

If a task moves from local work into web research, a cloud frontier model, connected-app use or browser automation, its cloud portion can use the plan's included usage or credits. That is why a credit-free local inference claim is not a promise of limitless free Computer activity. Claude's shared usage quotas are not denominated in equivalent Perplexity Computer credits, so the two plans cannot be compared by a universal per-task dollar figure.

Privacy settings and security controls

For Free, Pro and Max accounts, Perplexity calls its model-training control AI data retention. It is enabled by default. Turning it off stops subsequent data collection for AI model training, but Perplexity says previously collected training data cannot be removed through that opt-out. Perplexity states Enterprise data is not used to train models and offers additional retention and administrative controls. None of those policy statements by itself verifies the exact network telemetry sent by the fully local Mac implementation.

Anthropic's consumer privacy policies govern training and retention according to account settings and exceptions. For cloud Cowork tasks, files Claude fetches and processes are part of the cloud workflow. Managed Team and Enterprise plans have different contractual terms and admin options; their owners can control whether Cowork runs in the cloud while leaving permitted local desktop execution in place.

Agent permissions are a separate risk. A malicious instruction hidden in a document or webpage can try to redirect an agent, whether its model inference is local or remote. Claude's safety guidance specifically warns about prompt injection and actions that are hard to reverse. Ask before sending correspondence, restrict connected folders and apps, and keep credentials, banking or unrelated client records out of reach. Choosely's broader AI assistant permissions guide explains why agent authority needs limits.

Three buying decisions that make the tradeoff clear

1. Your business keeps sensitive project documents on a high-memory Apple silicon Mac. If a cloud AI model may not process those documents, Perplexity fully local is the relevant architecture to evaluate first. Keep it on local-only tasks, restrict access and verify organizational approval. The company claims local inference and agent work stay on-device; Choosely has not independently certified data egress or effectiveness.

2. You have a 24 GB or 32 GB Mac and routinely combine private notes with public market research. Perplexity Hybrid Compute has a documented route for this hardware. It starts in the cloud, delegates selected sensitive steps locally and uses a privacy gate. Review exactly what is permitted to leave. A 32 GB Mac does not automatically disqualify you from testing fully local mode, because 36 GB is a recommendation, but do not assume it will work well.

3. You want recurring, cloud-connected reports when your Mac is shut. Either Perplexity cloud Computer or Claude cloud Cowork can be considered. Compare required connectors, existing subscription, model preference, approval controls and usage limits. Choose Claude when you already rely on Claude workflows and accept Anthropic processing of the materials it opens. For Mac-only files, remember its desktop access conditions; a closed desktop app cannot supply newly requested local files.

The Choosely verdict

Perplexity fully local brings a genuine new choice for private-file processing on supported Macs. Its greatest practical difference from the current Claude Pro/Max Cowork route is that local model inference, agent planning and execution can take place on the device. A user who cannot send a client's file to an external AI model should investigate that route, while checking permissions, hardware and external calls carefully.

When cloud AI processing is acceptable, the architecture alone does not decide between brands. Perplexity Computer has a cloud route, and Claude Cowork does too. The sensible choice then comes down to the tools each can reach, how its usage is metered, and which workflow an individual or organization is prepared to authorize.

The remaining unknowns are meaningful: precise fully local compatibility and model storage, the extent of Mac-local session telemetry, the billing treatment of other local models and Pro's weekly allowance, and relative task completion quality. None was independently tested for this article. The source-backed comparison supports a conditional buying decision without manufacturing a performance winner.

Frequently asked questions

Does Claude Cowork still run locally on Mac?

On Pro and Max, new Cowork tasks run in the cloud from October 6, 2026, with no setting to switch new tasks back. Older local tasks and scheduled tasks remain exceptions. Team and Enterprise organizations can retain local desktop execution through administrative settings. Even those local execution routes use remotely served Claude models; Claude Code's data-flow documentation explains that local operations still send requests to a model provider.

What memory does Perplexity fully local need?

Perplexity recommends 36 GB or more unified memory for fully local Computer but has not published a distinct hard minimum. Its Hybrid Compute has a documented 24 GB minimum and 32 GB recommendation, on Apple silicon with macOS 15+. Find your memory under Apple menu → About This Mac.

Can Perplexity Computer run entirely offline?

Local inference can happen on a Mac, and reports describe offline tasks. Perplexity's announcement does not establish that every setup, update, app connection or workflow works without an internet connection. Treat offline suitability as task-specific until tested.

Do Perplexity local Mac tasks consume credits?

Perplexity says PPLX Pro tasks running locally do not consume Computer credits in its October 20 official announcement. That does not establish the credit rule for every local model or for cloud-escalated steps, or settle the effect on Pro's weekly included allowance. Choosely has not checked live account balances after a task.

Can Claude open a Mac folder from a task started on my phone?

Under Anthropic's current cloud-task documentation, the desktop-folder bridge requires the desktop app to be open and the task to have started on desktop or be a scheduled task. You can steer eligible cloud sessions from other devices, but a new phone-initiated task does not gain local-folder access merely because the Mac app exists.