# Use the documentation in your editor (MCP)

Fonte: https://lunahia.com.br/en/docs/mcp

The Luna documentation has a public MCP server. Connected to your editor or assistant, the AI searches the guides and the reference on its own, and reads each route’s exact contract before writing the integration code. The address is `https://lunahia.com.br/mcp`.

The server is read-only, asks for no key and stores nothing about you. It reads from the same content that generates this documentation, each page’s `.md` file and `llms-full.txt`: when a guide or the [API reference](https://lunahia.com.br/en/docs/reference) changes, the MCP changes with it, with nothing for you to update. It does not call the Luna API and does not send messages; for that use the [API key](https://lunahia.com.br/en/docs/account-and-keys) and the reference.

## The four tools

| Tool | What it does |
| --- | --- |
| `list_pages` | Lists the pages (guides and reference) with the identifier (`slug`), the title and a one-line summary. |
| `read_page` | Returns a page’s Markdown, identical to its `.md` file. Long pages come in pieces; the answer says which `offset` to ask for to continue. |
| `search_docs` | Searches the guides, routes, webhook events and error codes, and returns the best sections with `slug`, anchor, link and an excerpt. |
| `get_endpoint` | Returns a route’s OpenAPI operation (`method` and `path`), with the schemas already resolved: the exact contract to generate the client code. Big routes accept `part` (`request` or `responses`) to bring only half. |

All of them accept `lang`: `pt` (default) or `en`. Each answer has at most 24,000 characters; beyond that the text is cut with a notice, and `read_page` and `get_endpoint` accept `offset` to ask for the rest.

## Connect your client

In all of them, the name `luna-docs` is just a label; change it if you like. There is no authentication header to set up.

### Claude Code

One line in the terminal:

```bash
claude mcp add --transport http luna-docs https://lunahia.com.br/mcp
```

By default the setup applies only to you, in the current project. With `--scope user` it applies to all your projects, and with `--scope project` the command writes the `.mcp.json` below at the repository root, for the whole team to use. Check it with `claude mcp list` or, inside Claude Code, with `/mcp`.

.mcp.json

```json
{
  "mcpServers": {
    "luna-docs": {
      "type": "http",
      "url": "https://lunahia.com.br/mcp"
    }
  }
}
```

### Cursor

Create `.cursor/mcp.json` in the project, or `~/.cursor/mcp.json` to apply to all of them:

.cursor/mcp.json

```json
{
  "mcpServers": {
    "luna-docs": {
      "url": "https://lunahia.com.br/mcp"
    }
  }
}
```

### VS Code

Create `.vscode/mcp.json`. Note that VS Code uses the `servers` key, not `mcpServers`. You can also use the **MCP: Add Server** command from the command palette, choose **HTTP** and paste the address.

.vscode/mcp.json

```json
{
  "servers": {
    "luna-docs": {
      "type": "http",
      "url": "https://lunahia.com.br/mcp"
    }
  }
}
```

### Claude (remote connector)

On claude.ai, go to **Customize > Connectors**, click **Add** and **Add custom connector**, give it a name (for example `Luna docs`) and paste `https://lunahia.com.br/mcp` as the address. Under authentication, choose **No sign in**. Team and Enterprise plans: the organization administrator adds it, in **Organization settings > Connectors**. Then, in each chat, turn the connector on with the **+** button.

### ChatGPT

ChatGPT accepts a remote MCP server through **developer mode**, in the apps and connectors settings: create an MCP app, paste `https://lunahia.com.br/mcp` as the address and choose **no authentication**. Then turn it on in the chat. These menus change place and name between versions and plans, and Helsen could not check the steps against OpenAI’s documentation, so follow what your ChatGPT shows.

> Other clients that speak MCP over HTTP work the same way: the address is all they need. Claude and ChatGPT connect from their cloud, so they only see the published server, never a `localhost`.

## What to ask the AI

- “Using the Luna documentation, write the Node code that sends an approved template.”
- “What is the exact contract of `POST /v1/messages`? Generate the TypeScript types.”
- “How do I check a Luna webhook signature? Show an example in Python.”
- “What does the `window_closed` error mean and what do I do?”

## Technical details

- **Streamable HTTP** transport, with no session: each call is a `POST` with one JSON-RPC message, and the answer is a JSON. `GET` and `DELETE` answer `405`.
- Protocol versions: 2025-03-26, 2025-06-18, 2025-11-25 (with `initialize`) and 2026-07-28 (no handshake, with `server/discover`).
- Open CORS, for browser clients, and no credentials: there is no cookie or key to protect.

To test with no client at all, call a tool directly:

```bash
curl -X POST https://lunahia.com.br/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search_docs","arguments":{"query":"enviar template"}}}'
```
