# AI agents

> Let an AI agent find, connect, check and diagnose domains through the hosted CustomDomain™ MCP server, with human-approved scopes and no DNS records as input.

Source: https://docs.customdomain.ai/docs/use-cases/ai-agents

An AI agent can do the domain work for you or for your users: look up where a
domain's DNS lives, start a connection, follow it to live, and say why it is
stuck. It works through the hosted CustomDomain™ MCP server, the same control
plane as the API behind a tool interface.

The CustomDomain™ MCP server's only endpoint is `https://mcp.customdomain.ai/mcp`
(registry name `ai.customdomain/mcp`).

> **Only mcp.customdomain.ai is the product**
>
> Other MCP endpoints on customdomain.ai subdomains belong to the company's
> internal back office. They are not part of CustomDomain™; do not connect an
> agent to them.

## When this fits

* An agent you use (Claude Code, Claude Desktop, Codex, Cursor, VS Code, ChatGPT
  or a custom MCP client) should connect and look after your domains.
* Your product runs an agent that connects domains on your users' behalf.
* A coding agent is writing your CustomDomain™ integration and needs the docs in
  a form it can read.

## How CustomDomain™ fits in

* **One endpoint, Streamable HTTP.** Connect the agent to
  `https://mcp.customdomain.ai/mcp`. In Claude Code:

  ```bash
  claude mcp add --transport http customdomain https://mcp.customdomain.ai/mcp
  ```

* **Access a person approves.** An agent signs in with OAuth (authorization code
  with PKCE), and a signed in owner of your workspace approves exactly which
  scopes it gets (`domains:read`, `domains:connect`, `domains:disconnect`,
  `domains:purchase`). The grant can be revoked at any time. An API key works too, but it carries the whole
  workspace's authority. See [Agent access](/docs/agents/overview).

* **Seventeen tools, and none takes DNS records.** Record values are computed on
  the server from vetted templates, which closes off prompt injection that ends
  in an arbitrary DNS write. `connect-domain` returns a link for the person to
  finish (a sign-in with their provider, one click, or the records to add) and the
  agent follows the connection until it is live. Purchases through MCP are off on
  the hosted service.

* **A skill that teaches the safe way.** The [agent skill](/docs/agents/skills)
  tells an agent to treat domain names and DNS values as data, to ask before
  disconnecting, and never to purchase or write records.

* **Docs an agent can read.** [`/docs/llms.txt`](/docs/llms.txt) lists every page
  with a description, and every page is served as Markdown at its own URL plus
  `.md`.

An agent grant cannot be given a [test mode](/docs/guides/test-mode) application,
so an agent works on your live application and real domains.

## What to read next

- [MCP: drive it from an agent](/docs/getting-started/mcp): Connect, the flow an agent should follow, and what a failed call answers.

- [MCP server reference](/docs/mcp/overview): Every tool, its input and the scope it needs.

- [Agent access](/docs/agents/overview): Human-approved, scoped and revocable access with OAuth.

- [Agent skill](/docs/agents/skills): Install the skill that teaches an agent to use the tools safely.