Mac Mini AI Agent Remote Access From iPhone

Published 2026-09-26 · Updated 2026-10-11

A Mac mini AI agent setup pairs a small always-on Mac with Ask on the box and Poket PC on your iPhone so you can watch your agent, send new work, and use schedules or voice without sitting at a desk.

Watch-your-agent workflow

You (iPhone)                Mac mini
     |                           |
     |  Ask: "Sum up this PDF"   |
     |-------------------------->|  Claude Code or Ollama decides
     |                           |  each step, within the reach you set
     |  Optional: open Screen    |
     |<--------------------------|  H.264 main display, 30 fps
     |  Tap or Trackpad to steer |
     |                           |
     |  Disconnect               |  Task keeps running on the Mac
     |                           |
     |  Schedule: weekdays 08:00 |
     |-------------------------->|  Runs on the Mac, no screen

Before each task you pick its reach in Ask:

  • Screen (the default): your screen and your notes, no other tools.
  • Quiet: your notes only; nothing on screen is opened, clicked, or looked at. On the Claude subscription, Quiet tasks can’t run, because Claude Code works from a screenshot; Mac helper 1.0.3 and later refuse them with a message (earlier versions ended them without an answer).
  • Tools: everything set up on the Mac, including any tool that can run commands or send mail. It is never the default; once picked, it stays selected until you leave Ask.

Use Screen in Poket PC when you need to see exactly what the agent touched. Stop in Ask ends a task.

Schedules

Schedules run a task on weekdays or every day at a set time, or every hour. Scheduled runs are always quiet: they work from the agent’s memory and your notes and never drive the screen. They run only while the Mac is awake and the helper is open; if the Mac was asleep or the helper was closed at the set time, the task runs once, late, at the helper’s next check, and further missed times are not made up. The Claude subscription can’t run schedules, because Claude Code works from a screenshot and a scheduled run takes none; Mac helper 1.0.3 and later keep such a schedule off and say why, and the Local model option (Ollama on the Mac) does not need a screenshot. The two daily reports (07:00 and 18:30) run only on a Mac with a notes folder set up for the assistant.

Mac mini baseline

Follow the headless Mac mini checklist: helper in Applications, permissions, login item if you want the box reachable after reboot, and pairing with the six-digit check. Keep the mini awake and logged in.

On the software side:

Goal Install on the mini
Claude subscription path Claude Code
Local models Ollama plus the models you choose; Screen tasks send screenshots, so pick a vision model

Then choose the model in the iPhone app’s Settings. Read AI that controls your Mac for provider details. There is no screen in 1.0 for entering an API key; adding your own key is planned.

Remote access without opening your network

The mini can sit on home Ethernet while you connect from LTE or office Wi-Fi. Works anywhere uses the encrypted relay; you do not set up a VPN or port forwarding for Poket PC.

Voice

Voice lets you dictate a task, and the result can be read back to you. The wake phrase “Hey Mac” is off by default; turn on Listen for “Hey Mac” in Settings to start a task hands-free while the app is open.

Security habits

  • Revoke a lost phone from the mini’s menu bar immediately.
  • Keep reach as narrow as the task allows; read security on prompt injection.
  • Tasks do not stop to ask for your approval mid-run in 1.0. The reach you pick before tapping Run is the limit.
  • With Claude Code, your prompts and the screenshots Ask takes go to Anthropic under your account. With Ollama, they stay on the mini.
  • There is no Face ID gate on session open in 1.0—keep the phone locked.

Claude Code specifically

Ask uses Claude Code as its model: it runs the CLI headless, allows it only the Read tool to look at the latest screenshot, and performs each step itself. Claude Code’s own coding tools are not handed your task. See run Claude Code from iPhone and Claude Code for how Ask maps to the CLI you already use on the mini.

Cost

Poket PC 1.0 is free with no ads and no in-app purchases. You pay Anthropic or your own compute for models. A Pro plan is planned for version 1.1 and is not on sale yet.

Apps

Mac helper: download. iPhone: the public TestFlight beta on beta; version 1.0 is in App Review, not on the App Store yet.

Frequently asked questions

Can I run a Mac mini as an AI agent and control it from my iPhone?

Yes. Install the helper, install Claude Code or Ollama on the mini, pick it in the iPhone app's Settings, and send tasks by typing or by voice. The Claude subscription can't run scheduled tasks: Claude Code works from a screenshot, and a scheduled run takes none. Mac helper 1.0.3 and later say so and keep such a schedule off; use the Local model option for schedules.

Does the agent keep working when I close the iPhone app?

Yes. A task runs on the Mac and keeps going after you disconnect; when you reconnect, the app catches up on what it did. The helper must keep running and the Mac must stay awake.

Do I need an API key in Poket PC?

No. There is no key screen in 1.0. Use Claude Code with your Claude subscription or Ollama locally. Support for adding your own API key is planned.

Is this a Mac mini AI server farm?

No. In 1.0 the iPhone app holds one paired Mac at a time, and pairing another replaces it. Each mini can run its own helper, but you switch between them by pairing again.

What Mac mini specs does Poket PC require?

macOS 14+ on Apple silicon or Intel. Model choice for Ollama or Claude Code is your decision; Poket PC does not publish hardware benchmarks.

Related

Try Poket PC