Perplexity's Computer agent, which uses your apps and files to finish tasks, can now run entirely on your Mac. The update came out on Oct 20, as reported by MacRumors and 9to5Mac. Perplexity's own blog post wouldn't load for us at press time, so the details below come from those two reports. "Perplexity Computer" drew about 530 posts in our X sample this week.
What changed
Until now, Computer leaned on the cloud. In the new local-first mode, per 9to5Mac, "the AI model and the agent directing the task" both run on the Mac:
- On-device: the model, task planning, reasoning and file processing
- Sandboxed: code and tool execution run in isolated sandbox environments, per MacRumors
- Offline: tasks can run without an internet connection
- Files stay local unless you approve sending them to the cloud
With permission, Computer can work with local files and Mac apps including Notes, Messages, Mail, Keynote and Numbers. Perplexity's example task: review recent messages and emails, find deadlines, and build an action list in Notes.
What you need
| Fully local | Hybrid Compute | |
|---|---|---|
| Models | Qwen 3.6 or PPLX Pro, on device | Cloud frontier models + a local model |
| Memory | 36GB+ recommended (Apple silicon) | as little as 24GB |
| Works offline | Yes | No |
| Plans | Pro, Max, Enterprise | Available in the same Mac app |
PPLX Pro is, per MacRumors, "a post-trained version of Qwen". You download the models through the app. Perplexity highlighted the new Mac mini as a good fit.
Wondering why 36GB? Local models need memory for their weights plus the context they're working on. Our guide to how much VRAM a local LLM needs explains the math, and the local AI hardware sizing sheet helps you check a specific Mac.
When it reaches the cloud
Local doesn't mean locked in. When a task needs current information, browsing or cloud-connected apps, you can authorize cloud access and decide when to spend credits. You can also call on one of more than 15 frontier models for harder problems, per MacRumors.
Why it matters
- Privacy. Messages, mail and documents are the most sensitive data an agent can touch. Keeping them on-device removes a big objection to letting an agent read them.
- Local models are good enough for real tasks. A mid-size open model running a desktop agent is a sign of how far local AI has come. See this week's Flint vs Ornith local coding models.
- The cost moves to hardware. Running locally saves cloud credits but needs a high-memory Mac. Memory prices are a story of their own, covered in why RAM is so expensive.
Caveats
- An agent with access to Messages and Mail can also send from them. Check what actions you're allowing, and keep approvals on.
- Local models are smaller than frontier cloud models. Expect them to handle routine tasks well and to need the cloud option for harder ones.
Want to run open models on your Mac yourself, without an agent? Start with Ollama for beginners.
