Getting started
You have downloaded the app and you have nobody to ask but the person who sent it to you. This page is the whole path from a .zip in your Downloads folder to a piece of work you can look at, in the order you will hit it.
Version 1.1.3. Apple silicon Mac (M1 or newer), macOS 13 or later. To check yours: Apple menu → About This Mac. Beside Chip it says Apple M1, M2 or later; a Mac that says Processor and Intel cannot open this app.
Stuck at any point, stop and write. A person reads it: evanwm24@gmail.comopen in your mail app
Once you are signed in there is also a Send feedback button at the bottom of the sidebar, beside Settings, which goes to the same place and can bring the conversation you are looking at with it.
When something goes wrong in the app, press Report a problem — at the bottom of the sidebar, on the line that says something failed, or in the Help menu. It puts together a short report on your Mac: which version and which Mac, how you are signed in, how many teams you have, what happened last, and the app's own wording of recent errors. You read all of it before anything leaves, and it never holds names, messages, files, email addresses or keys. Signed in with an invite, press Send; on your own key, press Copy and paste it into an email to us. It answers the first questions we would ask you, so we can help sooner.
1. First open
- Unzip it and move it. Open the .zip from your Downloads folder (Safari usually does this for you, so Agent Platform may be there already), then drag Agent Platform into your Applications folder and open it from there.
- Let macOS open it. The app isn't signed by Apple yet, so the first time macOS says it can't check it. Close that message with Done (or Cancel, on macOS 13 and 14) — not Move to Trash. Then open System Settings → Privacy & Security, scroll down, and click Open Anyway next to Agent Platform. The button is there for about an hour after you tried to open the app; if it isn't, open the app again first. macOS asks once more: choose Open Anyway (Open on macOS 13 and 14) and, if asked, enter your Mac's password. You won't be asked again.
- Sign in. The app opens on a card headed Your AI team, working on your Mac. Type the invite code you were sent, your name and your email, and press Create my account. No AI account or key is needed — that is what the code is for.
Two smaller doors on that same card, in case you need them. I have my own Anthropic key takes you to the key box instead, if you already pay Anthropic directly and would rather the work went through your own account. Not now skips it; you can sign in later from Settings.
If the code is refused, the message says why rather than showing you a status number. A code has an end date: if it has passed, ask whoever sent it for a new one. Anything else, send that sentence to us as it appears — it is the useful part.
2. Set these three things up before anything else
The app will let you start work without any of this. It will just be work done by agents who know nothing about you, held only by the limits a new install starts with: 2,500 credits a day for the whole app, and 500 for any one task.
A team
You talk to your Executive Assistant. It answers you and does small jobs itself, and hands bigger work to a team. For work that is a team's, the app offers ready-made ones — research, writing, planning and a few more. On Today, where the app opens, they are at the very bottom, folded under Start from a ready-made team: click it, and choose a team from the list. The one you choose says what it does, and who is on it and what each of them does. Press Create, beside the list, when you have the one closest to your work. ＋ New team, at the bottom of the sidebar's team list, starts an empty team of your own at any time, and has the same list of ready-made teams under it.
Pick one when you have work like that. You are not choosing a structure for all your work, you are giving the assistant somewhere to send a team's work.
About you
Settings → About you. A plain text box headed About you and your work. Your Executive Assistant and your teams' agents read it before they work, unless a team's rule keeps it from the AI company that agent runs on, and your Executive Assistant and each team's lead add to it as they learn about you. Three or four sentences — what you do, who you work with, how you like things written — is the single cheapest improvement to everything that comes out.
If you would rather not write it yourself, the same screen has a box that says Or tell your Executive Assistant what to change, and it makes the change in its own chat.
Your usage limits
Settings → Usage, under Limits. See the next section for what each figure does. Set them to suit you before the first real run, not after.
3. The first thing worth trying
Open Today — at the top of the sidebar, just under Pause all work, and where the app opens. At the top of it is a box that starts a task; type what you want and press Start a task. Your Executive Assistant answers directly or hands it to a team, and the result comes back into that conversation. That is the whole loop, and it is worth doing once just to see the shape of it.
Then do the thing this app is actually for, which is a job you repeat:
- Open Settings at the bottom of the sidebar and choose Procedures, then press Start from a template… (＋ New procedure, beside it, starts a blank one instead).
- Choose Chase what you still need from the Template list. It reads your notes from a first call or meeting, lists everything you still need from someone before work can start — files, logins, decisions — drafts one short chaser for each, and then stops and waits for you.
- Pick the Team that will do it and a Budget per run. The template suggests 300 credits.
- Press Create, then Run, and paste in your notes from a real call or meeting when it asks.
The point of choosing this one first is the last step. It is an approval: "Read the chasers. Edit anything you like, then approve. Nothing is sent from here — you send them yourself." You get to see the whole thing work against your own material with nothing at stake.
There are two other templates on that list when you want them: Pressure-test a recommendation, which looks at a draft recommendation as the owner paying for it, the staff who would live with it and a sceptic testing every claim; and Checked deliverable, where a second team on a different AI provider checks the first team's work before it reaches you. That last one needs two teams, and the app will not let a team check its own work.
One honest limit on it: an install signed in with an invite code only sees Claude models. Reaching another AI company's models needs a gateway key, which is not something you can add from a settings screen — ask us and we will set it up with you. Until then the check step will correctly refuse to let a team check work done on the same provider, so that template is the one thing on this list you cannot finish on your own.
4. What a rehearsal is, and why to use it first
When you run a procedure, the dialog has a tick box:
Rehearse — do the real work and show you the result, but change nothing: no files written, nothing sent.
It is ticked for you the first time you run any procedure, and it is the reason to trust the second run. A rehearsal is not a preview or a description: every step does the real thinking against the real material and produces the real result — the actual list, the actual drafts — with every tool that writes a file, runs a command or sends a message removed. Where a step would have acted, it says what it would have done.
Two things to know before you use it.
- A rehearsal costs the same as a real run. The app says so in the run dialog. It is the same work; only the writing and sending are gone.
- It is marked afterwards. A rehearsed run is labelled Rehearsal everywhere it appears, and approving one finishes the rehearsal rather than doing the thing for real. To do it for real, run the procedure again without the tick.
If you connect a mailbox, use a Gmail account made for your agents, not your own. Connect Gmail, in Settings → Email, signs in with Google: if Google ends that sign-in (its access removed, the account's password changed, or a long time unused), the app stops checking mail and asks you to sign in again. A Google app password, under Advanced, needs no sign-in with Google. Settings → Email walks you through both.
Email has its own version of this, if you connect a mailbox. Settings → Email has a button marked Preview these rules, which shows you what your rules would have done with the mail you actually received over the last 7, 14 or 30 days. That one is free — it reads senders, subjects and dates only, nothing starts, nothing is sent, and it does not move the watcher's place in your inbox. The Rehearse reply button beside an individual message is a paid model call, and the app labels it as one.
5. What the usage limits do
Everything the app spends on AI is counted in credits. One credit is one cent of AI cost at the AI company's list price, so a task that used 40 credits cost 40 cents of model usage. The figures in the app are what the work actually cost, so they are worth watching even while somebody else is paying.
There are four kinds of usage limit, all in credits:
| Limit | What it does |
|---|---|
| Daily limit | Work stops for the day once it has spent this many credits — the backstop against a runaway loop. Requests already under way still finish and are counted, so a day can end over by what they cost: on a large model with several copies at once, that can be as much as the limit itself. A new install starts at 2,500. A team or a team member can be given a lower daily limit of its own. Your Executive Assistant has one of 500: past it, it does no work itself until the next day, and a short hand-off of your message to a team still goes through. 0 means no limit. |
| Most per task | Each chat message, task, procedure run, routine run and email task stops once it has spent this, all of its work together. Requests its agents have under way then still finish and are counted, so a task can end over by what they cost: on a large model with several copies at once, that can be as much as the limit itself. A new install starts at 500. Each can be given a lower figure of its own, never a higher one. 0 means no limit. |
| Monthly credits | A team's credits for the month, set in that team's own limits. When they are used up, that team's work, and its sub-teams', stops, and every other team keeps working. |
| Included credits | While your invite pays for the model usage, the credits it includes. Settings → General shows how many are used. A request sets aside about 100 while it runs, so about 100 fewer than these can be spent. |
The first two, and a Time limit per task (minutes) — a task that is still going after this is stopped; 0 means no limit — are under Settings → Usage → Limits, in the app's own words above.
A procedure has a budget of its own as well, set when you create it and shown in the run dialog: Stops at 300 credits, and waits for you wherever it needs your approval.
Two things about how these behave, because they are the difference between a ceiling and a number on a screen. They are checked while work is running, not between tasks — so a single long job cannot sail past one. And they count money that has been committed but not yet billed, which matters when several agents are working at the same time: an earlier version of this app measured $3.80 spent against a $1 ceiling for exactly that reason, and that is what the counting fixed.
They are not exact to the cent. A job that is already running finishes the model call it is in the middle of, so expect a small overshoot rather than a hard stop on the number.
6. When something is refused
This app refuses things on purpose, and a refusal is meant to be readable. The sentence you are shown names the cause and what to do about it. Send us the sentence, not a summary of it.
"Everything is stopped until the day rolls over, or until you raise the limit and resume"
The daily limit. It names the limit and what has been spent today, in credits. Work starts again by itself when the day rolls over; to go on today, raise it in Settings → Usage and press Resume. If you did not expect to be near it, look at Today first — something has been running that you did not mean to run.
All work is paused
Pause all work sits at the top of the sidebar and stops everything that is running, starting nothing new until you resume. When it is on, the sidebar says All work is paused with a Resume button, and the box on Today tells you the same thing. Nothing is exempt from it — not a routine, not a procedure halfway through. If work is refused for no reason you can see, check there first.
A routine that did not run
A routine that was refused says why, on the routine itself, instead of sitting there looking like it will run tomorrow. Two ordinary causes: all work is paused, or the team has nobody accountable for it — see below.
"The … team has no sponsor, so it cannot run on its own"
A team can have a named person — a sponsor — accountable for it. If that person is removed and nobody replaces them, the team stops doing anything unattended: no routines, no triggers, no procedures. You can still sit down and run that team's work yourself, because then you are the named person. The message tells you where to fix it: name a sponsor for the team in Settings → People.
An agent says it could not reach a website
Agents read only the sites they are allowed to. Settings → What it can do → Reading the web holds the list, and the same screen has Where they have been: every site your agents reached for, and every one they were stopped from reaching. A team or an individual member can narrow that list further, never widen it.
It stopped and is waiting for you
Not a refusal. Today → Needs your input is where everything waiting lands, including Procedures waiting for you. Nothing moves past an approval until a person answers it, and no agent has any way to answer one on your behalf.
Require Touch ID for approvals, under Settings → Pre-approvals, is off unless you turn it on. Off, the approvals that send, spend or widen something are one click: another AI you have allowed to operate your screen could make them too, and the Record marks each one “without Touch ID”. Turning it on asks for nothing while the Record can be written (while it can't, it needs your Touch ID, as every approval then does); turning it off asks for your Touch ID, so that AI can't switch it off.
With it on, on a Mac with Touch ID, those approvals ask for your finger. On a Mac without it — one with no Touch ID sensor or Touch ID keyboard, one where Touch ID is not set up, or a MacBook whose lid has stayed closed since the app opened — they are one click, and so are the changes in Settings that Touch ID would guard: another AI you have allowed to operate your screen could make them too. The app says so where you approve and at the top of Settings until you press Got it, and always under Settings → Pre-approvals. That tab lists a pre-approval only while one is in force, so you can see it and revoke it; no screen in the app makes one. It never asks you to type a confirmation instead: anything that can click can type.
"Your invite account has been switched off"
The access your invite gave has ended or been paused. The request it stopped was not sent or charged; anything done before it is kept. Write to whoever sent your invite, or add your own Anthropic key under Settings → General to keep working on your own account.
"Your invite sign-in is no longer accepted"
The service no longer takes this Mac's sign-in. Your first code has been used, so it cannot sign you in again: ask whoever sent your invite for a new code for your email, then press Sign in again under Settings → General. Or add your own Anthropic key there.
"You've used your … included credits"
The model usage on your invite comes with included credits, and they have been used. Ask for more; Settings → General shows what has been used. Until the access service is updated it may say this the older way, in the access service’s own words, from before it was updated: "You've reached your beta spending limit", with the figure in dollars. It means the same.
Two things to know that are not in the app
The Mac has to be awake. Routines and procedures run while Agent Platform is running on a Mac that is awake and plugged in — from 0.7.2 that includes with its window closed, when its icon is in the menu bar. Closing the lid sleeps the Mac and the work stops. Today has a card, Keep working while you're away, that holds the machine awake until a time you set and tells you afterwards whether it managed it — but it cannot keep a closed laptop awake, and it says so. The one exception is a procedure: Settings → Procedures → Worker lets an Admin name another computer that runs procedures while this Mac is closed. Agent Platform does not come with one, so for most people work stops when the Mac does.
Nothing is sent to anybody without you. Email is off until you turn it on, per agent, with your own limits. The built-in procedures stop at approved drafts and you send them. If you want to check that for yourself before trusting it, that is what a rehearsal is for.
Updates
When a new version is out, the app tells you: a line on Today, and Install update in Settings → General → App version. Press it and the app downloads the new version with your sign-in and checks it was signed by us; nothing closes. When it's ready, press Restart to install and it opens again as the new version. It won't while something is working or waiting for you. The old version goes to the Trash once the new one has started — and if the new one doesn't start, the app offers to go back to it; your chats, teams and saved work stay exactly as they are. Using your own Anthropic key instead of an invite? The same place offers Download from the site: replace the app in Applications. You can also ask it: Check for updates, beside the version. What changed in each version is on the What's new page.