Welcome to Aitroop

Think of Aitroop as a teammate with their own computer. You describe the work; they design the workflow, run it in a sandboxed environment, return typed outputs you can read, share, or chain into the next step. You don't write code — you have a conversation, and the conversation becomes a saved App you can run again, schedule, hand to a colleague, or call from an API.

Reading order. If you have 15 minutes, jump straight to the Quickstart — it walks you end-to-end, with screenshots and exact messages to type. Come back here when you want the conceptual map.

Unlike traditional chatbots

Most AI tools are a single chat box. You type, it replies, you move on. The next morning your teammate asks for the same thing and re-types the prompt from scratch — losing your tweaks, your context, your refinements. Aitroop works differently:

What you'd do in a chatbotWhat you do in Aitroop
Re-type your "company briefing" prompt every Monday.Save the prompt as a one-form App. Click Run. Or schedule it for 9 AM Monday.
Copy the model's output, paste it into Notion, format it, send it via email.Declare the output as a typed Artifact. Aitroop delivers it formatted, to email/Slack/Drive.
Worry about whether the agent will read your Gmail correctly.Authorize a Connect once. The agent uses your scoped OAuth token transparently.
Hope the model doesn't make stuff up about your private data.Every run executes in an isolated Sandbox with file access, network, and the right tools.
Hand-coordinate "research, then human review, then draft email" across three threads.Define a 3-stage App: agent → human → agent. The platform handles the pause-and-resume.

The one-paragraph version

Open a chat. Say "Create an app that…". The agent activates its App Builder skill, asks 1–5 clarifying questions, designs the workflow (form fields, stages, output format, required Skills and Connects), validates everything, saves it via POST /api/apps, and tells you the App ID. You then run the App from your library — fill the form, click Run, get a typed Artifact. Or attach a Schedule and it runs automatically every Monday at 9.

What you can actually do with this

Three concrete scenarios, each runnable today:

📰 Weekly intelligence briefing

Build a "Competitor Weekly" App that takes a list of company names. Every Monday at 9 AM, it searches the web for last week's news on each, writes a 2-page Markdown brief with sections per company, and emails it to your team's distribution list. You touched the prompt once; the platform runs it 52 times a year.

📧 Inbox triage

Connect Gmail (read-only) once. Build a "Daily Triage" App that reads the last 24 hours of unread mail, ranks the top 20 by urgency with a one-line summary each, and posts it to a Slack channel every morning at 8. Authorize Gmail-compose later when you want it to draft replies for you.

🗂 Multi-stage outreach pipeline

A 4-stage App: (1) find companies matching an ICP description, (2) enrich each with funding history, (3) pause for you to remove rows you don't want, (4) draft personalized cold emails for the survivors. The human pause is built into the App — the run waits indefinitely until you approve, then resumes from stage 4.

Where to start

1
Quickstart — 15 minutes, end-to-end Sign in, build a Company Brief App by chatting with the agent, run it on Stripe and Notion, schedule it. Read this first.
2
Core concepts — the schema in detail The App JSON document. All 12 input types. The 3 stage types. The {{field_id}} template syntax. Reference companion to the Quickstart.
3
Chat and Apps — the two surfaces Chat is the cockpit. Apps are the saved workflows. Most of your time lives in these two pages.
4
Deep dives, when you need them Skills · Connects · Executions · Schedules · Sandbox · Artifacts.

What an App looks like under the hood

You don't write this — the agent does — but seeing the shape makes everything else click:

{
  "name": "Company Brief",
  "icon": "🏢",
  "input_schema": [
    { "id": "company_name", "type": "text", "required": true },
    { "id": "tone", "type": "select",
      "options": [{ "label": "Executive", "value": "exec" }, ...] }
  ],
  "stages": [
    { "id": "research", "stage_type": "agent",
      "goal": "Research {{company_name}} and write a brief
          in a {{tone}} tone...",
      "timeout_ms": 600000,
      "artifact_defs": [
        { "id": "brief", "format": "markdown" }
      ] }
  ],
  "resources": { "skills": ["web-search"], "connects": [] }
}

The four primitives

Master these four and the platform is small enough to hold in your head:

PrimitiveOne-line definitionLives where
App recipeA JSON document describing a runnable workflow.app_def table · Sidebar → Apps
Execution runOne trip through the App. Status, stage logs, artifacts.app_execution · App's Runs tab
Artifact outputTyped deliverable a stage produces. Markdown, CSV, code, image, file.app_artifact · stored in S3
Connect credentialAn OAuth bridge to your account on another service.app_user_connect · Settings → Connects

Two surfaces, one platform

Almost everything you do is in one of two surfaces — and the right surface depends on the repeat question:

SurfaceUse whenPersistence
ChatYou're exploring. One-shot tasks. Casual questions. Debugging a failed run.Conversation history. Sandbox is destroyed on idle.
AppYou'd want to run this task again. A teammate should be able to do it without your prompt know-how. You want it scheduled.Forever. Versioned. Auditable. Runnable from API.

A productive chat almost always wants to become an App. The transition is one button: Save as App, or just say "save this as an app".

How Aitroop differs from a Notebook / Workflow tool

It's tempting to compare Aitroop to Zapier, n8n, or Jupyter. The closest analogy is none of those:

ToolBuilding unitHow a "step" is configured
Zapier / n8nNodes wired in a graphClick through forms per node
Jupyter / NotebooksCode cellsYou write Python / R
ChatGPT / Claude.aiConversation threadRe-prompt every time
AitroopAn App with stagesYou describe the stage in English; the agent runs it inside a real sandbox

The key word is real. An Aitroop stage isn't a prompt template — it's an agent that reasons, calls Skills, hits Connects, executes code, retries on failure, and produces a typed artifact. The stage's English prompt is the spec; the agent's behavior is the implementation.

If you get stuck

The agent itself is the fastest help. Open a chat and ask "How do I…?" — it has the App Builder skill loaded and knows every part of the platform. For specifics:

  • Your first App fails — open the failed run, click Debug in chat. See Executions.
  • Connect won't authorize — usually a scope issue or wrong account. See Connects.
  • Schedule isn't firing — workspace time zone or paused schedule. See Schedules.
  • Anything elsetalk to us.