How it works

Four steps, and the fourth one is you.

Ace is not a model and does not host one. It is the part that turns a sentence into a plan, drives the runtime on your own machine, and knows when to stop and ask.

The loop

Say it, route it, run it, approve it.

  1. 01

    You state an outcome

    Hold the key and say what you want to be true when it is done — not the steps, the outcome. “Get me ready for Thursday.” Ace runs in the menu bar, so you do this without leaving what you are looking at.

  2. 02

    Gold Agent plans and routes it

    The objective is turned into a plan and handed to the mode that fits: Notes for capture, Red Agent for anything long-running, Partner for a goal with a date on it, Trader for a chart. The planning reasoning runs through your own AI account.

  3. 03

    It runs on your files

    The runtime — bundled Codex, or Claude Code — operates on the real files on your Mac. This is the step a chat box cannot do: Ace reads the files where they live and sends the request and selected context directly to the AI provider account you connected.

  4. 04

    It stops and shows you

    The result comes back as something finished — a file, a plan, a set of notes — and waits. Nothing is sent, filed, published or treated as done until you approve it. There is no code path that skips this.

Where the work happens

Which side of the line each mode sits on.

The distinction that matters is not "cloud or local" — it is what crosses the wire and to whom. Here it is per mode.

Gold Agent
Plans through your AI account. The planning seat. One objective in; Ace plans it, routes the work, and brings the result back for review.
Notes
Recording stays on this Mac; the transcript is summarized by your AI account. Captures meetings and classes locally, then sends the transcript text to your connected ChatGPT or Claude account to reduce it to structured notes and answers.
Red Agent
Runs in the background. Takes the long-running task so the front of your day stays free, and reports its own state.
Partner
Off until you switch it on. Keeps your goals on track. It is a card you turn on yourself; off until you do.
Stealth
On this Mac. Keeps private coursework private on the machine in front of you. Not a way to hide anything from anyone entitled to see it.
Trader
No order is ever placed. Reads the chart you show it and gives disciplined market context. Signals only: execution is off.

Questions

About the mechanism.

Is Ace a model?

No. Ace has no model of its own and does not host, train or fine-tune one. It plans, routes and drives — the reasoning happens through the account you connect.

What is the difference between this and using Claude Code directly?

Claude Code is a runtime you drive from a terminal, one task at a time, with you supplying the context and the sequencing. Ace is the layer that takes a spoken outcome, decides what the steps are, picks the mode, runs them, and returns a finished thing for approval — without you leaving the app you were in.

Does it work offline?

The application runs locally, but reasoning needs your AI provider, so a request needs a connection. If it cannot reach your account it says so out loud rather than failing silently.

What if it gets something wrong?

You see the result before anything happens with it, which is the point of stopping at approval. It is a drafting and execution assistant, not an oracle — read what it hands you.

See it on your own Mac.

Try for free Download first

Free to start: your first 20 tasks and 5 minutes of Partner. No card, no charge, nothing to cancel. Make your account and Ace downloads from your account. Privacy.