Perplexity Routes Sensitive Computer Tasks to Local Macs With a Privacy Gate

Perplexity has added hybrid compute to its Computer agent platform, letting a cloud model hand confidential portions of a task to a local model on an Apple silicon Mac. The pitch is a single workflow that can use cloud models for research and planning while keeping private files and device data local.

The feature is available now in Perplexity’s desktop app for opted-in enterprise customers and Pro and Max subscribers with an Apple silicon Mac running macOS 15 or later. Windows and Linux versions are planned for later.

One workflow, two starting points

Perplexity says its cloud orchestrator breaks a request into subtasks, then routes web research, long-horizon planning and heavy reasoning to cloud models. Work involving private files, local data or actions on the device can go to a subagent running on the Mac, with the task retaining its context across the handoff.

That is the inverse of Perplexity’s local-first Computer version for Nvidia DGX Spark, launched the previous week. That product starts work on the user’s device and escalates a step to cloud models only with permission; hybrid compute starts in the cloud and delegates downward when a task touches sensitive material.

The classifier decides when privacy controls apply

Before information goes to the cloud, Perplexity’s on-device Privacy Gate scans for personally identifiable information, including names, addresses, account numbers and secrets. When it flags material, the user decides whether that portion should run locally or be shared with the cloud.

That design offers a practical control, but it concentrates risk in the classifier. Perplexity says users can expand and inspect what the gate flagged before transmission; a false negative would still allow sensitive information to reach the cloud.

Controls for the organization and the machine

For enterprises, administrators can set one organization-wide sensitivity policy and audit data leaving each device. Perplexity says locally generated tokens do not use cloud credits, though its credits still cover orchestration and delegation.

The local tradeoffs

• Users can choose Google’s Gemma E4B, Alibaba’s Qwen3.6 35B-A3B, or Perplexity’s post-trained Qwen3.6 35B for local inference.
• Perplexity recommends at least 32GB of unified memory for its better local-model tier and says its smallest option significantly underperforms the larger Qwen models.
• macOS’s Seatbelt sandbox constrains local-agent actions; Perplexity says an attempted disallowed action stops and requests user permission.
• Perplexity does not currently permit unrestricted YOLO-mode execution for local agents.