> ## Documentation Index
> Fetch the complete documentation index at: https://docs.arlet.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Create with chat

> How to create an MCP server by describing it in a conversation

Just describe the tool you want, and Arlet handles everything from implementation to deployment. No code required.

There are two ways in: **inside a conversation with a connected AI service**, and **in the chat screen on the Dashboard**. Both use the same generation pipeline, so the resulting server is identical.

## Contents

* [Create by talking to your AI service](#create-by-talking-to-your-ai-service)
  * [Editing from the conversation](#editing-from-the-conversation)
* [Create in the Dashboard chat](#create-in-the-dashboard-chat)
* [After creating](#after-creating)

## Create by talking to your AI service

With Arlet connected to an AI service such as ChatGPT or Claude, you can build a server right inside that conversation. There is no screen to switch to, and the tool you create is usable from the same conversation.

### How to create

1. Connect Arlet to the AI service you use (see the pages under [Connect Arlet](/connect/claude) if you haven't yet)
2. Ask for the tool you want in the conversation
   * Example: "Build a tool that fetches the weather forecast for a given city from a weather API"
3. The AI confirms the specification one question at a time. It asks only about **what you want** — what the tool should do, which service and which data it works with, which operations you need. Technical parts such as authentication and configuration are filled in by Arlet
4. You're offered 2–3 ways to build it; once you pick one, a short spec summary is presented. Approving it starts the implementation
5. If a key for an external service is needed, an input panel opens inside the conversation, along with guidance on where to obtain the key (you never paste keys into the chat)
6. Once deployment completes, you can call the tool from that same conversation

<Note>
  If you belong to more than one workspace, you'll be asked where to create the
  server. You are not asked about the implementation language — your default
  from User Settings is used (Python if unset).
</Note>

### Editing from the conversation

After deployment, ask for a change in the same conversation and the implementation is updated. Changes are presented as a summary of **what will change** rather than as code, and deploy once you approve. If something goes wrong, you can roll back to the previous version.

## Create in the Dashboard chat

Use the Dashboard chat when you'd rather see the generated server's details and required settings on screen as you go.

### How to create

1. Select **Create with chat** in the server list on the [Dashboard](https://app.arlet.dev/)
2. Describe the tool you want in natural language — the API to integrate with, the tools (features) you need, and what the inputs and outputs should look like. Refine the specification over the course of the conversation
3. When the specification is settled, choose **Build from this**, review the implementation instructions assembled from the conversation, and select **Run build**
4. When generation finishes, the panel on the right shows the server's details and its **required environment variables**. Register keys for external services there
5. Once every required environment variable is set, **Deploy** becomes available. After deploying, **Open project** takes you to the hosting screen

## After creating

* **Editing**: whichever way you created it, you can keep refining the server from either the AI service conversation or the Dashboard chat
* **Environment variables**: keys for external APIs can be changed later from the server's settings screen
* **Editing as code**: use **Export to GitHub** in the Dashboard settings, then simply push your changes and they deploy automatically. To use a repository you already have, see [Deploy a repository](/how-to-use/deploy-repository)

<Note>
  Created servers become available through the
  gateway — no reconnecting and no refreshing of the
  tool list required.
</Note>