Perplexity's Computer Agent Now Runs Fully Private on Apple Silicon Macs

Perplexity has added a fully local mode for its Computer agent on Apple silicon Macs. According to the launch announcement, model inference, planning, tool orchestration, and file processing can run on the Mac without consuming Computer credits or sending task data to Perplexity’s cloud.

The release follows two earlier configurations. Personal Computer gave the agent access to local files and Mac apps. Hybrid Compute kept orchestration on the device while routing selected work to cloud models. Fully Local keeps the complete agent loop on the Mac.

The agent loop moves onto the Mac

Computer combines a language model with an orchestration layer that interprets a goal, chooses tools, reads permitted data, and executes each step. In Fully Local mode, those components use the Mac’s processors and unified memory.

With macOS permission, the agent can read and modify files and work across Notes, Messages, Mail, and Numbers. Perplexity’s example pulls tasks from Notes, checks a budget request in Messages, compares planned and actual figures, drafts a status update, and creates a meeting invitation.

Local inference does not prevent requested actions from reaching other services. Messages, email, calendar invitations, and connected apps still send data through their respective providers when the task requires it.

The model list needs verification

Perplexity’s published materials identify different model versions. The announcement names Qwen 3.6 and PPLX Pro, a post-trained Qwen model, while a related company post identifies Qwen 3.8 as the base for PPLX Pro. The Hybrid Compute page lists Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model.

Those labels do not align, so teams evaluating the product should confirm the exact model, version, license, and memory requirement shown in the current app. Models download through an in-app picker, removing the need to install Ollama or configure a separate inference runtime.

Hybrid mode redacts before upload

Hybrid mode keeps the orchestrator on the Mac and routes selected requests to more than 15 cloud models for advanced reasoning, live web information, browser tasks, or connected services.

An on-device personally identifiable information classifier scans each outbound request. It replaces detected names, email addresses, phone numbers, and account numbers with placeholders, then restores the original values after the cloud response returns. Perplexity has released the classifier through its Secure Intelligence Institute.

The classifier’s protection extends only to information it detects. Indirect identifiers, proprietary text, and unsupported data formats can still reach a cloud model, making representative testing necessary for regulated or confidential workloads.

Sandboxing contains tool execution

Code and tool execution occur in isolated sandboxes with controlled access to files and connected apps. Perplexity says the local product uses the same isolation design as its cloud-hosted Computer agent.

Sandboxing limits a process’s reach, while granted permissions determine which files and apps the agent can change. A deployment review should cover folder scope, app entitlements, connector credentials, outbound actions, audit logs, model updates, and the handling of deleted or overwritten files.

Bounded workflows fit best

Perplexity highlights workflows involving private documents and repeatable operations:

• Portfolio review: Compare private holdings with investment guidelines and downloaded market data, flag concentration risk, and update an Excel workbook and Keynote deck.
• Client proposal: Convert confidential discovery notes, a client brief, and an internal rate sheet into a scoped proposal, budget workbook, and Outlook draft.
• Morning triage: Scan Messages and Mail, extract requests and deadlines, create an action list in Notes, and draft follow-up replies.

The listed Qwen3.6 35B-A3B model uses a mixture-of-experts architecture, which activates a subset of its roughly 35 billion parameters for each token. That design makes laptop inference more practical, although difficult research and multi-step web tasks will generally receive stronger reasoning from large frontier cloud models. Fully Local suits private, bounded work with limited need for current web data; Hybrid mode covers tasks that need cloud reasoning or external context.

Memory determines eligibility

Local Computer is available to Perplexity Pro, Max, and Enterprise subscribers through the Mac app. Compatibility depends on the selected model and available unified memory. Perplexity’s cited materials do not specify each model’s download size, storage requirement, or performance across individual M-series chips.

Three steps to start

1. Install or update the Perplexity Mac app on a supported Apple silicon Mac.
2. Download a compatible model from the in-app selector.
3. Choose Fully Local when starting a Computer task.

Administrators evaluating a broader rollout will also need to account for model storage, download distribution, endpoint permissions, thermal limits, and update policies.

The deployment boundary shifts

On-device planning and execution can keep private files out of Perplexity’s inference infrastructure while reducing variable cloud costs. The packaged Mac app also replaces several components that teams would otherwise assemble themselves, including a local model runtime, agent framework, app integrations, and sandboxed executor.

The architecture transfers operational responsibility to the endpoint fleet. Model-version clarity, permission controls, auditability, and measured accuracy on real tasks will determine whether Fully Local can satisfy an organization’s security and reliability requirements.