# MCP server

Spamgone has a hosted MCP server, so AI agents such as Claude, Cursor and VS Code can check messages and manage your setup: rules, keys, settings, the review queue, metrics, logs and usage.

The server is at `https://spamgone.dev/mcp`, over Streamable HTTP. Connect with an API key in the `Authorization` header. The agent gets exactly that key's rights.

## Set up

Claude Code:

```bash
claude mcp add --transport http spamgone https://spamgone.dev/mcp --header "Authorization: Bearer $SPAMGONE_API_KEY"
```

Cursor, in `.cursor/mcp.json`:

```json
{
  "mcpServers": {
    "spamgone": {
      "url": "https://spamgone.dev/mcp",
      "headers": { "Authorization": "Bearer ${env:SPAMGONE_API_KEY}" }
    }
  }
}
```

VS Code, in `.vscode/mcp.json`:

```json
{
  "servers": {
    "spamgone": {
      "type": "http",
      "url": "https://spamgone.dev/mcp",
      "headers": { "Authorization": "Bearer ${env:SPAMGONE_API_KEY}" }
    }
  }
}
```

Other clients that speak Streamable HTTP work the same way: the URL plus the header.

## Options

Add these to the URL:

| Option           | What it does                                                                                                           |
| ---------------- | ---------------------------------------------------------------------------------------------------------------------- |
| `read_only=true` | Lists only tools that change nothing.                                                                                  |
| `features=...`   | Lists only the named tool groups: `checks`, `review`, `metrics`, `logs`, `keys`, `rules`, `usage`, `settings`, `docs`. |
| `team=...`       | Refuses the connection unless the key belongs to this team.                                                            |

For example: `https://spamgone.dev/mcp?read_only=true&features=checks,metrics`.

## Safety

- A check-only key lists only `check_message`, `give_feedback` and `search_docs`.
- Check results never include the checked text, so instructions hidden in a spam message never reach the agent. When an agent asks for text with `include_text`, it comes back marked as untrusted user content.
- Tools that revoke keys or delete rules do nothing until the user confirms: the agent must call them again with `confirm: true`.
- Tools take only their own fields. No tool accepts questions or instructions for the AI model.
- Tool calls count toward your rate limit and quota like API calls, and show in Logs with the source `mcp`.
