A hired gun for the Mac you can’t get to.

Write the job. Send a one-time code. They paste one line into Terminal, and an AI agent does the work on their Mac while you watch. Then it reports back and leaves.

Nothing for them to sign up for, and no API key. macOS only for now. You can set a code to work once, and you can stop the agent from your dashboard at any time.

Example: a job ticket with its one-time code, and what the person at the Mac sees in Terminal after pasting the command.
JobCheck Maya’s Mac is ready for day one
Works in~/Desktop
K7QM-4XPA

Example code. Real ones are made in your dashboard.

Agents are for people who live in Terminal. Your dad lives in Mail.

About half of U.S. adults now use AI chatbots1. But chatting isn’t the same as an agent that opens a folder and gets something done. Running one today still means a terminal, API keys, installs and config, so most people never try.

Agents today

Built for people who live in a terminal

Pick a tool, install it, sign up, generate an API key, paste it in the right place, then write a good prompt. Every step is a place to give up.

  • install runtime
  • create account
  • get API key
  • edit config
  • write prompt
With Hired Gun

The expert sets it up. They paste one line.

You know what needs doing, so you write the job and hold the keys. The person at the computer needs only Terminal and a code.

$ npx gigagent K7QM-4XPA

1. Pew Research Center, About half of U.S. adults now use AI chatbots, up from a third in 2024 (survey conducted Feb 17–23, 2026). The statement about agents is our own observation, not a survey figure.

From “can you take a look?” to done, in three steps.

No screen share. No remote desktop. No back and forth.

  1. Write the job

    In the dashboard, say what needs doing in plain English. Pick the one folder it may work in, choose a model, and decide when the code stops working.

  2. Send the code

    You get a one-time code. Text it, email it, read it over the phone. They open Terminal and run:

    npx gigagent K7QM-4XPA
  3. Get the report

    The agent does the work and returns a plain-English summary for them, and the answer, key findings and changes for you. Then it removes itself from their Mac.

You stay in control the whole time.

  • Watch liveEvery command and file the agent touches shows up on your dashboard as it happens.
  • Send instructionsSpotted something mid-job? Add an instruction and the agent picks it up.
  • Stop anytimeStop a session or revoke the code. The agent’s next model call is refused and it halts.

The jobs you’d normally do over someone’s shoulder.

Anything you could do by looking around a Mac and tidying or writing something down, a person can hand to an agent with one line.

  • IT teams

    New-hire onboarding and laptop readiness

    Replace the 45-minute setup call with a checklist the new hire kicks off themselves.

    “Check this Mac is ready for day one: OS, updates, disk, FileVault, required apps.”

    You get a pass/fail readiness report before IT spends a minute on the laptop.

  • Remote IT and helpdesk

    Diagnostics for “why is my Mac slow?”

    The agent gathers the facts you’d ask for over the phone, in the order you’d ask.

    “Find out why this Mac is slow. Check disk, memory and top processes.”

    You get disk, memory and the top offenders, plus a likely cause.

  • Family tech support

    Desktop clean-up and disk space rescue

    Skip the 40-minute call. The agent is instructed to move files rather than delete them, and to log every move so it can be undone.

    “Sort the Desktop into folders by type and tell Dad what is taking up space.”

    You get a tidy Desktop, a plain-English guide and an undo log.

  • MSPs and agencies

    Setting up client machines

    Send the same job to every client. Each one runs a single line and you get a comparable report.

    “Audit this client Mac and list what is installed, what is out of date and what is missing.”

    You get consistent audits across clients without a site visit.

  • Schools and labs

    Student and lab machine checks

    Hand a code to whoever is at the machine and collect results without visiting every desk.

    “Report free disk space and macOS version on this lab Mac.”

    You get a fleet snapshot gathered by the people sitting there.

  • Small businesses

    IT help without an IT department

    Whoever is the unofficial computer person writes the job; staff only paste a line.

    “Check backups, disk space and system updates on the front-desk Mac.”

    You get routine health checks without hiring or a contractor visit.

  • Support and asset teams

    Info for tickets and asset inventory

    Stop asking “can you go to About This Mac and read me the number?” and get it in one pass.

    “Collect macOS version, FileVault status, serial number and installed apps.”

    You get a structured record you can paste straight into a ticket or sheet.

  • Anyone with a messy folder

    File organisation projects

    Works inside the one folder you choose. Duplicates are flagged and moved aside rather than erased.

    “Group Downloads into Documents, Images, Installers and Other. Flag duplicates.”

    You get sorted folders, a list of duplicates and the undo log.

Honest about v0. The agent is instructed to stay inside the folder you chose, not to install software and not to use admin rights. Those are rules it follows, not locks: the commands it runs are not sandboxed, so only hand a code to someone you’d trust with that, and watch the live view. It is not a replacement for a full device-management tool. It runs on macOS only for now, and needs Node.js 22+ on the Mac.

See what an hour of setup call is really costing you.

Every onboarding call burns two people’s time: IT and the new hire. Plug in your numbers and see how much of it comes back.

Typical laptop setup and onboarding help takes 40 to 60 minutes of hands-on IT time.

Default 2: the IT person and the new hire. Both their time counts.

Optional: support tickets

Ticket time is counted for one person (the support agent) to stay conservative.

Estimated time returned each year

417 hours
$37,483saved per year
14.4work-weeks returned

hours = (hires × min per hire × people ÷ 60) + (tickets/mo × 12 × min per ticket ÷ 60)
dollars = hours × hourly cost · work-weeks = hours ÷ 40

Example at scale: 1,000 hires × 50 min × 2 people = 100,000 minutes ≈ 1,667 hours ≈ 42 work-weeks a year, before counting a single support ticket.

These figures are estimates based on the inputs you enter. They are not guarantees, and real savings depend on how many jobs you hand to Hired Gun and how long they take today. Model usage costs are not subtracted; they are typically cents per job.

The report comes back structured and ready to use.

No scrolling through a chat log. When the job finishes, the agent hands in a report: a plain-English summary for the person at the computer, and the answer, key findings and changes for you.

  • Summary in plain language for the person at the keyboard.
  • Answer and key findings for you, easy to paste into a ticket.
  • Changes listed one by one, so nothing is a surprise.
Email (optional) Slack, Discord or webhook (optional) Always on your dashboard
Laptop readiness: Maya’s MacBookExample report. Finished in 2m 14s
Done

Summary for Maya

Your laptop is nearly ready. Everything important is set up. One thing is left: a macOS update, which we will schedule for tonight.

Answer

Ready for day one, with one fix needed: macOS is two versions behind.

Key findings

CheckResultStatus
macOSTwo versions behindUpdate
FileVaultOnPass
Free disk61 GB of 245 GBPass
Required apps6 of 6 presentPass

Changes

  • wrote laptop-readiness.md
  • moved nothing

Made to be handed to someone who isn’t technical.

Some protections are built into the system. Others are rules the agent is told to follow. We keep the two apart so you know which is which.

Built into the system

  • The AI key never reaches their Mac

    The server proxies every model call. Their machine only ever holds a short-lived session token, never your provider key.

  • Revoke a code, stop the agent

    Revoke a code or stop a session and its next model call is refused. The agent has no model access without the server’s say-so.

  • One-use codes

    A one-use code works once; run it again and it’s rejected. You can also set a code to last until the job succeeds, or for a set number of hours.

  • You watch every command live

    Every command the agent runs, its output (up to 2,000 characters each) and the files it opens stream to your dashboard, with the option to send instructions or stop.

  • Cleans up after itself

    When it exits normally, it removes its temporary files from the Mac. A power loss or force-quit can leave them behind.

Rules the agent follows

These are instructions given to the agent, not technical locks. Its commands are not sandboxed, so treat them as strong guidance rather than a guarantee.

  • Works in one folder

    You choose the working folder. The agent is told to change things only inside it. It may run read-only commands elsewhere to look things up, such as the macOS version or installed apps.

  • Moves instead of deleting

    The agent is told not to delete files, empty the Trash or overwrite anything. Anything that has to go is moved aside.

  • Logs every move for undo

    The agent is told to write each move to an undo log in the folder, so a clean-up can be reversed.

  • No sudo, installs or network

    The agent is told never to use admin rights, install software, or download or upload anything.

Questions people ask before they send a code.

What do they need?
A Mac with Node.js 22 or newer, plus Terminal. Node.js is a free installer from nodejs.org, and IT-managed Macs often have it already. If it’s missing or too old, Hired Gun explains what to do. (If Node.js isn’t installed at all, macOS just says command not found: npx.) They don’t need an account or an API key.
Does it install anything permanently?
No permanent install. npx fetches the helper on demand, it runs once, and it removes its temporary files when it exits normally. A power loss or force-quit can leave them behind. The agent is also told not to install software.
What can it see?
It’s told to work inside the one folder you pick for the job and to change nothing outside it, though it may run read-only commands elsewhere to look up information such as the macOS version or installed apps. Its commands, their output (up to 2,000 characters each) and the files it opens are streamed to you, so you see what it sees. macOS may ask the person to allow Terminal to access that folder, which is normal.
What if it goes wrong?
You can stop it from your dashboard at any time, and the person can press Ctrl+C. The agent is told to move files rather than delete them and to write each move to an undo log in the folder, but that is a rule it follows, not a technical lock. Revoking the code cuts off its access to the model immediately.
Which models does it use?
Models run through Vercel AI Gateway, so you can pick from Claude, GPT, DeepSeek and others per job. The model call goes through the server, never directly from the recipient’s Mac.
Who sees the commands and files?
You do, as the operator: the dashboard shows each command, its output (up to 2,000 characters) and the files the agent opens. That same content is also sent to the model provider, through Vercel AI Gateway, so the model can do the job. Don’t point it at a folder whose contents you wouldn’t want you or the provider to see.
Does it work on Windows?
Not yet. Hired Gun is macOS only today. Windows and Linux aren’t supported in this version.
What does it cost?
The cost is model usage. For a typical job like a folder clean-up or a system check that is usually cents per job. The exact amount depends on the model you choose and how long the job runs.
Does the person need an account or API key?
No. They need the code, Terminal and Node.js 22 or newer. Everything else, including the model access, is handled on your side.
Why does the command say “gigagent”?
Hired Gun was called GigAgent until October 2026. The command keeps its old name for now so that codes, notes and links people already have keep working. Renaming the package is a separate step that hasn’t been done.

Stop driving over. Send a code.

Write the job, share the code, and watch it get done. Your first one takes about a minute.