# Getting started

> Create an agent, hand it a PDF and watch it download it to its machine, read it and answer with the exact figures. Ten minutes, no code.

URL: https://www.ghosty.studio/en/docs/getting-started

**When you finish** you will have an agent with its own machine —disk, terminal, memory— that has already read a document of yours and answered with exact dates and amounts. Every step shows the real screen you will see.

## Before you start

An account at [ghosty.studio/signup](/signup) (Google, Apple or email). Nothing else.

| | |
|---|---|
| **Agent** | Name, identity and engine. It is yours; it works wherever you activate it. |
| **Machine** | Its disk and terminal. Sleeps on its own, wakes on the first message. |
| **Channel** | How you talk to it: the Studio chat, Teams, your editor or your software. |

## 1. Create the agent

In **Agents** press **+ Nuevo agente** (New agent). Three doors:

![New agent: Describir, Plantilla, En blanco](/docs-img/nuevo-agente.png)

- **Describir** (describe): write in one or two sentences what it will do ("answer my hardware store's customers on WhatsApp, quote and schedule deliveries") and Studio proposes a name and identity.
- **Plantilla** (template): pick one of the eight [house recipes](/en/docs/getting-started/recipes) (support, WhatsApp sales, appointments, collections, content, HR, team assistant, dev in the repo) or drop your own `.yaml` recipe.
- **En blanco** (blank): name and engine, nothing else.

![Templates: eight recipes and the drop zone for a .yaml recipe](/docs-img/plantillas.png)

All three fill the same form: name, what it is for, identity (editable), workspace, engine and model. For this guide pick the engine **Ghosty · Lite**: it is the one with chat inside Studio. (Claude and DeepSeek are used from Teams and WhatsApp; they have no "Hablar" button in the list.) Press **Crear**.

Studio provisions its machine in a few seconds. Once the agent shows up in the list with **Hablar** (Talk), it is ready.

![Agent row in the list with the Hablar button](/docs-img/hablar.png)

## 2. Hand it a document

Press **Hablar**. In the chat, attach a PDF with the clip (or drag it in) and type:

```text
Summarize this document in three points and tell me which dates and amounts appear.
```

![Chat composer with a PDF attached](/docs-img/adjuntar.png)

## 3. Watch what it does

The first time takes a few seconds: it says **"despertando la caja"** (waking the machine) while it boots. Then you see every tool it uses, in order. With this message it is two terminal commands: download the file and read it with `pdf-reader`.

![Two Terminal calls, marked as completed](/docs-img/herramientas.png)

And the answer: not a vague summary, but the dates and amounts that are in the document.

![Agent reply with three key points, exact dates and amounts](/docs-img/respuesta.png)

You can close the tab mid-turn: the work runs on its machine and the answer is there when you come back.

## 4. Choose where else it works

In **Agents → your agent** is its card: model, identity and channels.

![Agent card: Teams and WhatsApp channels, and the card to configure it from your coding agent](/docs-img/canales.png)

- **Ghosty Teams** — flip the switch and give it an `@handle` in [Teams](/en/docs/channels/teams): it reads the channels where it is mentioned and works with documents and forms.
- **Your coding agent** — under **Configura a Ghosty desde tu agente de código → Generar token** you get the skill command and the token; from there Claude Code or Cursor set its identity, files, skills and MCPs: [Configure your agent](/en/docs/configure).
- **Your editor** — the same block includes the Zed or VS Code configuration; the same agent, same machine, over [ACP](/en/docs/protocols/acp).
- **Your software** — the [API](/en/docs/api/authentication) with OAuth2.
- **WhatsApp** — turn on the channel and tick the number it serves; the number is connected once from Sales. Guide: [WhatsApp](/en/docs/channels/whatsapp).

## What just happened

- An **isolated machine** was created just for your agent.
- The PDF did not travel inside the message: it was uploaded to your library and the agent **downloaded it to its disk** through a signed URL that expires.
- The agent **chose the tools** (terminal, `pdf-reader`) and ran them there; you saw every step.

## Next steps

- [Concepts](/en/docs/concepts): turns, permissions, connectors, skills, engines.
- [Teams](/en/docs/channels/teams): put it to work with your team.
- Ask for something bigger: "Make a one-page document with this summary" returns a `.docx`.
- [Security](/en/docs/reference/security): what isolates your agent and where your keys live.