The user guide

How to use AgentsPoppy

Every screen you'll meet, from download to your first running poppy. The app is free, the one-time AWS setup takes about three minutes, and nothing below ever asks for — or uses — admin access to your cloud.

Building apps instead? That path lives in the developer hub.

Install it and open it

Download AgentsPoppy for macOS, Windows, or Linux — it's free, with no account to create. The first launch introduces the one idea everything else rests on: AgentsPoppy is the gatekeeper between your apps and your own AWS. Apps never hold your keys; they get scoped, short-lived credentials that expire within the hour, and you can switch any of it off whenever you like.

First launch. The green pill is a promise the app is built around: it never asks for or uses admin access.
Choose your cloud

Poppies run in your cloud account, so setup begins by choosing one. Today that is Amazon Web Services — Google Cloud and Azure are marked coming soon — and the screen is built for people who have never opened a cloud console: one choice, full screen, nothing else to parse. No AWS account yet? Press the AWS mark and the app opens Amazon's own signup; creating an account is free, new accounts get free credits, and plenty of the Free Tier stays free permanently. You'll enter a card for identity verification; Free-Tier usage isn't charged.

Step one of four. On a first run there is deliberately no way out of the setup but through it — a newcomer can't fall into an empty app by accident.

The two screens Amazon shows a brand-new user: the Free Tier page, then the signup form. Ten minutes, start to finish — then press I already have an account.
Give AgentsPoppy its own key-holder

Two minutes in the AWS console: you make one dedicated, least-privilege user — never your account root. One action per card, in order: a button that opens the right console page, and a button that copies the scoped policy. That policy lets the one-time setup create AgentsPoppy's own role and operator, and nothing else, and you can revoke it any time. Stuck on any card? A small Stuck? link under it explains that step in plain words — the help is there when needed and invisible when not.

Step two. The policy travels by copy button rather than a link, so this step can't break if a URL ever moves.
Take the keys to your cloud

Create that user's access key and paste both values in — the only thing you type, and the last console work you'll ever do here. The wizard walks you to AWS's own Done button on purpose: the secret is shown only once, and leaving that screen half-finished is how people lose it. If your company signs in through SSO and can't create access keys, an advanced path is offered right on this screen, exactly where you'd need it.

Step three. Your key is used once, in memory, to create AgentsPoppy's own limited operator — it is never written to disk and never uploaded.
Pick where it lives — and let it finish itself

Last question: which region of the world should hold everything AgentsPoppy sets up? The closest one to you is suggested from your timezone — closest is fastest, and your data stays in that part of the world. Choosing a flag arms the finish button, and pressing it does the rest on its own: AgentsPoppy creates the broker role and a non-admin operator in your account, links the account, and verifies the connection actually works before saying so. This is the part that makes AgentsPoppy different — the role lives in your account, AgentsPoppy only ever assumes it, and it never receives a key of its own. Safe to walk away from: if it's interrupted nothing elevated is saved, and running setup again picks up wherever AWS actually got to.

Step four. Regions are shown by flag and place, not just eu-west-1, with the closest suggested — and the finish button is the confirm, so the choice can never be skipped.

Done — verified, not merely assumed. AgentsPoppy proved it can assume the role before celebrating.
Install a poppy from the catalogue

The Poppies tab is the app store — except every listing links its open repository, so you (or your AI agent) can read exactly what a poppy does before it ever touches your cloud. Pick one, press Install, done.

Every card carries “Read the source”. Developers can also share poppies outside the catalogue — installing from your own disk always stays open.

The first time a poppy wants to use its connection, AgentsPoppy stops and asks you — Approve or Deny, in plain language, with every requested power rated green, amber, or red. Nothing runs until you say yes, and you can pause or revoke a poppy's access at any time.
Use it like any other app

A running poppy is just an app — inbox, dashboard, buttons — with one difference you can see in its title bar: Sandboxed · own resources only. Its backend lives in your account, its data stays in your account, and where costs exist they're counted in front of you.

CrewPoppy — AI agents living in the user's own AWS, with a live spend meter and hard monthly limits.
See everything it built — and tear it down in one click

Open any poppy on the Dashboard and you get the view no ordinary app store can give you: a live map of what it actually built in your cloud, service by service, plus every power it holds — each one labelled “Its own” (scoped to resources it created) or “Broad” (and then read-only). Done with it? “Tear down everything it built” removes the lot. No zombie servers, no surprise bill.

MailPoppy's footprint: Lambda, DynamoDB, Cognito, S3, SES — every node clickable, every permission disclosed, and the undo button right at the top.
Audit updates before they run — with your own AI

This is the part no other app platform does. When an update ships, AgentsPoppy doesn't auto-install it — it shows you Review and update. Reviewing reads the open source; nothing is downloaded to your computer until you choose to install.

One click on “Verify this update with your AI agent” copies a prompt — written by AgentsPoppy, not by the app being audited — that has your own AI read the source and report anything undeclared or risky: new external calls, credential access, broader AWS powers.

The audit running in an AI coding agent. Note what the prompt does: it tells the agent to treat everything the update says about itself as untrusted claims to verify — the update can't sweet-talk its own auditor.

The verdict, in your agent's own words. Only now do you decide whether to press install.

// why this matters
Every other store asks you to trust an update because someone else reviewed it. Here the reviewer works for you, reads the actual source, and answers to nobody else — and if the verdict is bad, the update simply never touches your machine or your cloud.