Everything Agent Mac Does
32 built-in tools, 7 AI providers, and the systems that hold it together. No marketing gloss — this is the spec.
New here? You don't need to understand any of this to start. It begins in a careful mode, asks before it changes anything, and explains itself as you go.
Build your own team of agents
Give the work a name. Pick a starting point from 6 ready-made role templates, set what it should take care of, and you have a working agent in two clicks. Make a few and you have a team in minutes — each one is yours to rename, re-scope and reshape.
Executive assistant
Keeps your calendar and inbox in shape and drafts what needs answering.
Research assistant
Digs into any topic on the web and leaves you a tidy brief.
Mail triage
Sorts, flags and drafts replies — never sends without you.
Briefing writer
A morning note: your day, your inbox, anything that moved overnight.
Librarian
Keeps a folder you choose tidy — renames, sorts and files what lands there.
Monitor
Watches the pages you name and tells you when something changes.
Its own everything
Each agent gets its own access, schedule, memory, skills and inbox — and only touches what you allow.
@mention or just ask
Type @name in chat to send a job straight to that agent, or ask the main agent to create, schedule, retune or retire one. Deleting an agent always asks you first, at every level.
Uploads, on your terms
Let an agent post images you approved to sites you named, with daily limits. Agent Mac holds every upload to your list, and an up-to-date Agent Browser double-checks it.
You set the leash
Five presets, from draft-everything to full autonomy, per agent, with an after-the-fact ledger of every risky thing it did. The two that run without asking (Complete mission and Bypass permissions) can send mail and messages on their own, so the app warns you and asks you to confirm before you switch one on.
Your Agents' Filing Cabinet
A shared cabinet
Folders of documents you and your agents have filed — Inbox, Reports, Receipts, whatever you make. Tagged, searchable, and yours on disk.
Agents file their work
When an agent produces a report or grabs a screenshot, it files it away — with a note of who filed it and where it came from. It can read a PDF page by page, and create folders when you ask.
Per-agent access
Each agent reaches only the folders you grant it, to read or to write. Nothing else in the cabinet is visible to it.
Outside content, flagged
Anything that comes from outside — a web page, an imported file or a scanned letter — is marked as outside content, so an agent asks before acting on what it says.
On your phone
Browse the whole cabinet from Agent Remote and scan paper documents straight into it. The phone can add scans but never moves or deletes a thing.
32 Built-in Tools
Most tools have their own on/off switch, and screen control, Messages and Mail start switched off. Autonomy levels decide what each one may do without asking.
agent
Starts sub-agents, each with its own context and tools, to work on parts of a task.
background_task
Runs a separate agent in the background while the conversation carries on.
browser
Drives a real Chromium browser through the Agent Browser app — navigates, clicks, fills forms, reads pages.
browser_fill_secure
Types a saved card detail into a page without the AI ever seeing it. Asks first.
browser_service
Starts, checks and self-heals the browser backend; pairs with Agent Browser.
calendar
Agent Mac’s own calendar — events show on your Mac and iPhone, badged when the agent added them.
claude_code
Hands big coding jobs to a headless Claude Code session. Needs Claude Code installed and signed in; runs bill to that account, normally your Claude subscription rather than API tokens. Asks first.
cloud_gpu
Rents and runs a cloud GPU for your model — start, stop, check cost. Every purchase needs your approval.
computer
Full screen control — sees, clicks, types and scrolls. Off by default, even at YOLO, and asks before it acts.
contacts
Looks people up in your Mac’s Contacts. Read-only.
file_delete
Moves files and folders to the Trash, so they can be restored.
file_read
Reads text files on your Mac. Protected files — private keys, keychains, credential and browser password stores, Agent Mac’s own data — are off-limits.
file_search
Finds files and folders by name inside a specific folder, like your Documents — not across your whole Mac at once.
file_write
Creates and edits files on disk.
filing_cabinet
A searchable library of the documents and screenshots your agents file for you.
generate_image
Creates images from a description, with your OpenAI or Gemini key.
grep
Fast content search across files and folders.
history_search
Searches your past chats and task runs.
Sends email through Apple Mail (only when Agent Mail isn’t installed). Off until you switch it on. Sending asks first while you’re there, unless it’s a scheduled task you set up or a Team agent you’ve set to run without asking.
memory_write
Long-term memory as editable markdown files on your disk.
messages
Sends and reads iMessage and SMS. Off until you switch it on. Sending asks first while you’re there, unless it’s a scheduled task you set up or a Team agent you’ve set to run without asking.
notes
Notes, tasks and checklists in Agent Mac’s Task Manager, on your Mac and iPhone.
pause_task
Pauses work that’s waiting on you and picks it up when you’re back.
reminders
Reads and manages your Apple Reminders.
scheduled_task
Full cron syntax with active-hours windows.
screen_text
Reads the text on your screen, on-device. Off by default, with computer control.
shell_command
Runs zsh commands with real output.
skill
Loads saved skills — reusable know-how for recurring jobs.
team
Creates, schedules and hands work to your Team agents.
view_image
Looks at images, PDFs and SVGs.
web_fetch
Fetches and reads web pages.
web_search
Searches the web for current information.
Works While You're Away
Give it a job once and it keeps doing it, quietly, while your Mac is on.
Schedules in plain language
Daily, weekly, monthly or one-time from a friendly editor, or full cron if you like it precise. Every task shows who runs it, Agent Mac or one of your Team agents, and you can reassign it with a click.
Active hours
Out of the box, runs start between 7am and 11pm. Widen it, narrow it or run round the clock in Settings ▸ Automation.
Heartbeat check-ins
Every 15, 30 or 60 minutes it looks through HEARTBEAT.md, your list of what to keep an eye on, so you come back to done, not to a to-do list.
Background tasks
Hand off a long job and carry on chatting. Up to three run at once, and if more than one needs the browser they take turns instead of fighting over it.
Careful when you’re not looking
Your Mac needs to be awake with Agent Mac running. Scheduled runs never use screen control or Claude Code, and they refuse to send or change things after reading outside content.
Missed runs
A run that was due while your Mac was asleep happens when it wakes. Every run is logged in the Activity window.
Notes, Tasks & Calendar
Its own Task Manager and calendar
Agent Mac keeps its own notes, tasks, checklists and calendar, on your Mac and your iPhone. Things your agents add are marked as theirs.
Quick Capture
Press ⌥⇧N from anywhere to jot down a note or task without opening anything. Prefer it always in view? A small floating bar does the same.
Apple Reminders and Calendar, if you want them
Two-way sync with an “Agent Mac” list in Apple Reminders and an “Agent Mac” calendar in Apple Calendar is optional. It stays off until you switch it on in Settings ▸ General ▸ Apple Sync.
Contacts and Reminders
It can look people up in your Contacts (read-only) and manage your Apple Reminders when you ask it to.
Pay Safely: the Card Vault
Let it buy things for you without ever handing the AI your card.
Sealed in the Secure Enclave
Cards are encrypted with a key that never leaves your Mac’s Secure Enclave, so a copied vault file is useless on any other machine. The AI never sees a card number: Agent Mac types it into the checkout itself.
Every payment is yours to approve
Each payment needs Touch ID or your password on the Mac, or Face ID or Touch ID on your iPhone, at every autonomy level. The amount is checked against the checkout page first.
Limits you set
Defaults are £100 a payment and £250 in any 24 hours. Lowering them is instant; raising them waits 24 hours, and they can never go above £500 a payment or £1,000 a day. Pounds only, for now.
Virtual cards only
It only pays with a virtual card, the kind a bank lets you cap and cancel, or a card with its own limit, and only on the sites you list for that card.
Any AI, Even One on Your Own Mac
Agent Mac isn't tied to one AI company. Bring a key, run a model yourself, or rent the horsepower.
Bring your own key
Use Claude, GPT, Gemini or Kimi with your own API key. Pick a model per chat, per scheduled task or per Team agent, and let Auto choose a lighter one for easy turns.
Models on your own Mac
Run open models with llama.cpp (one click installs it and a small starter model, about 2 GB) or Ollama, with no key and no per-use cost. Apple’s on-device model works too, through a free helper called apfel. Your prompts stay on your Mac, and it keeps working with no internet.
A cloud GPU, set up for you
Want a bigger open model than your Mac can run? Say “set it up for me” and Agent Mac rents a cloud GPU, starts the model server and connects to it. RunPod, Vast.ai, Lambda and DataCrunch are driven directly; any other cloud connects by web address or an encrypted tunnel.
You approve the price
It stops for exactly three things: your sign-in, your approval of the live price with a spending cap, and connecting. Renting or ending a GPU only happens with your one-time approval, at every level, and an idle GPU is stopped. The GPU is billed by that company, never by us.
MCP Client
Connect any MCP server and its tools appear alongside the built-ins. Ready-made connectors include GitHub, Notion, Canva, Slack, Linear and Sentry (you sign in to each), and local ones such as filesystem, git, fetch and Playwright. Add your own, or request one. Agent Browser's 39-tool server registers automatically when Agent Browser is installed, and its tools replace the built-in browser tools while it's connected. Claude Code, Claude Desktop, Cursor, Windsurf or another MCP client on your Mac can also hand tasks toAgent Mac over a one-click MCP bridge. Those runs are capped at the lower of your own autonomy level and the level you set for remote runs, and they hit the same approval gates as phone runs: gated actions go back to Claude Code to approve, and a desktop app's need your approval on the Mac.
Eco Mode
Per-role provider and model routing to cut API spend: cheap models for cheap roles, with a preference for the Claude Code handoff when Claude Code is signed in with your Claude subscription, so work runs on that instead of API tokens. Includes a screenshot-describe economy option.
Memory
A typed memory store with background consolidation — the agent files away what it learns while idle. Memory is markdown on your disk: open it, edit it, delete it. Conversations live in local SQLite. None of it is stored on our servers.
Instruction Systems
Skills
Reusable know-how the agent loads when a task calls for it. Write one, have AI draft it for you to review, or add one from a link (it’s checked and shown to you first, and starts switched off). Each Team agent can have skills of its own.
Playbooks
HEARTBEAT.md defines what background check-ins should look at.
AGENT.md cascading
Per-folder instructions that cascade down your directory tree.
Hooks
Run your own commands at points in the agent lifecycle.
Context compaction
Long sessions compress automatically instead of failing.
Session resume
Pick up any previous session where it left off.
Activity Window
Every run is logged: plan → tools → results → reflection, with duration and token count. Each run has a plain-English “Why?” so you always know what the agent did and what it cost.
Voice Conversation — just talk to your Mac
Press ⌥⇧Space anywhere and talk. Voice Conversation runs the real agent from your voice — it reads and writes files, runs commands, and browses the web, speaks the answer back, then re-opens the mic for what's next. Approvals are spoken: say approve, deny, or always allow, behind the same permission gates as typing. Rich output — tables, code, long text — shows on a floating console while it reads you the headline. Two voice engines: OpenAI Realtime, the recommended one with the most natural back-and-forth, billed per minute on your own OpenAI key; or the on-device engine, which needs no OpenAI key and has no per-minute charge (the agent's replies still bill your AI provider as usual). If you add an OpenAI key and have never picked an engine, Agent Mac switches to Realtime and tells you. If Realtime can't connect, the on-device engine takes over. Change it any time in Settings ▸ Voice.
Prefer to type? ⌥Space opens the command palette, and Dictation transcribes straight into the palette, chat, or Quick Capture. Any reply can be read aloud. Every spoken conversation is saved to your history, readable afterwards like any chat. Voice works from your iPhone too — see Remote.
New to AI? It Teaches You as You Go
You shouldn't need a manual, or a friend who knows computers. Agent Mac is built to explain itself.
A guided start
Setup takes a few short steps: choose a cloud key, a model on your Mac or a rented GPU, and bring your history across from ChatGPT, Claude or Gemini if you like.
A guide in the Help menu
Take the Tour, browse the Agent Mac Guide, or press ⌥⌘/ on any screen and choose “Explain This Screen”. Hints and tips can be switched off whenever you like.
Careful by default
It starts on Smart: it does safe reading and searching by itself, and asks before it changes files, deletes things or runs most commands.
Plain English everywhere
Every run has a “Why?” in everyday words. On your iPhone, notifications, approvals and reports are written in plain English too, not tool names.
A permissions check-up
Settings ▸ Permissions shows which macOS permissions Agent Mac has and which it needs, and what each is for.
Experts are not left out
Cron schedules, hooks, AGENT.md, Skills, Team agents, any MCP server, local models and cloud GPUs are all there, and go as deep as you like.
Shortcuts, Accessibility & Language
Keyboard shortcuts
- ⌥Space
- Open Agent Mac from anywhere (the command palette)
- ⌥⇧Space
- Start a Voice Conversation
- ⌥⇧N
- Quick Capture: jot down a note or task
- ⌥⌘/
- Explain This Screen
- ⇧⌘C
- Copy the last response
- ⌥⇧⌘C
- Copy the whole conversation
Accessibility
VoiceOver labels throughout, and Reduce Motion is respected.
Language
The app is in English for now. With OpenAI Realtime, voice can be pinned to any of 20 languages.
Everyday touches
A Copy button on every reply, copy a whole chat in one click, rename any chat, and the chat you just used moves to the top.
Self-Healing Storage
If the database ever fails to open, a recovery database catches the session, and a conservative, atomic importer merges it back later — additive-only, never destructive. Your history survives bad days.
Move In, and Never Lose It
Bring your history
Import your conversations from ChatGPT, Claude and Gemini — and plain transcript files. They stay searchable, and you can pick any of them up where you left off.
The agent fetches it
Rather than hunting for an export, ask the agent to fetch it — or drop the file in Downloads and Agent Mac offers to import it.
Scheduled backups
Back up on a schedule to your own iCloud automatically, or to a folder you choose.
Restore onto a new Mac
Moving to a new Mac? Restore everything — conversations, memory, settings — from your latest backup on the first screen.
Keys stay out
API keys are left out of backups unless you deliberately add a passphrase, so a backup is never a place your secrets leak.
Platform
Native Swift 6
A real menu bar app — not a wrapped web page.
macOS 14 Sonoma or later
Intel and Apple Silicon.
Notarized & signed
Notarized by Apple, hardened runtime, Developer ID.
7 AI providers
Claude Fable, Opus, Sonnet & Haiku, the GPT-6 family, Gemini, Kimi, any Ollama model, open models through llama.cpp on your Mac, a cloud GPU or your own server, and Apple's on-device model (macOS 26 or later on Apple Silicon, through apfel, a free helper you install once). Cloud model lists ship with each app update; Ollama and llama.cpp list the models your own server reports.
21-day trial
No account, no card.
Lifetime license £44.99
One payment. No subscription.
iCloud — Three Honest Roles
Communicate
Lock-screen notifications from your Mac reach Agent Remote through your own iCloud. They carry short labels only, like which tool needs approval, and never what’s said. Turn on Away access in Agent Remote and, when your phone can’t reach your Mac directly, messages go through your iCloud too, sealed end-to-end. There’s no company server, so we never see any of it.
Back up
Every day by default, to your own iCloud Drive when you’re signed in to it (otherwise to this Mac). You can pick a folder instead, or switch on “iCloud — automatic”. We never see your backups, and API keys and saved logins stay out unless you add a passphrase.
Share cloud GPU keys
Only if you switch it on: your cloud GPU keys reach your other Macs through iCloud Keychain.
Calendar, notes and tasks reach Agent Remote over the same connection as chat (your home network, or your iCloud when you're away). There's no separate iCloud sync. Pairing survives an Agent Remote reinstall because your iPhone keeps it in its own Keychain, not in iCloud.
Try it free for 21 days.
Download it, use it properly, then decide. If it earns a place on your Mac it's one payment of £44.99, never a subscription.
Download Free Trial- Free for 21 days
- No account
- No credit card
- Notarized by Apple