---
url: /handbook/governance/request-from-your-ai.md
description: >-
  Ask for an agent from the AI chat you already use. Approval happens in
  DutyGraph.
---

# Request from your AI

An employee can ask for an agent from their own AI chat. The chat talks to DutyGraph over the Model Context Protocol (MCP).

## The tools

| Tool | What it does |
| --- | --- |
| `get_my_holdings` | Lists the access DutyGraph has on record for you |
| `validate_agent` | Previews an ask against your holdings and the rules. Writes nothing. |
| `create_agent_request` | Creates your request in DutyGraph and submits it to its approval chain. If you name a scope no agent may hold, it creates nothing. |
| `get_request_status` | Shows where your request is in the approval chain. Only the owner or requester can read it. |

The tools take "me" from your token, never from an input. A person recorded as having left cannot use them. In AI apps that support MCP Apps, the proposal shows as a read-only card. Elsewhere it shows as a table.

::: warning
**Approval happens in DutyGraph, not in this chat.** No MCP tool can approve, change or deny a request.
:::

## Which Task Cards a chat request uses

A request needs Task Cards. If your chat names them, DutyGraph uses those. If not, DutyGraph tries to match your words to a proposal already set up for you. If it cannot pick one, it asks the chat to name the Task Cards. It never invents a Task Card from your words.

The tools apply the same evaluator as the app. They check the catalog, your holdings and the rules. They do not check that your Task Cards need each scope, because rule CEIL-02 is not enforced yet. See [The key join](the-key-join.md).

## Connect

You need a personal token with only the `request_agents` scope. In DutyGraph, open **Agents › Requests › Connect your AI**.

**Claude Code**

```bash
claude mcp add --transport http dutygraph <origin>/api/mcp \
  --header "Authorization: Bearer <token>"
```

**Claude Desktop**

Claude Desktop needs the `mcp-remote` bridge. Follow the steps in **Agents › Requests › Connect your AI**.

## Example

> "I reconcile vendor invoices every Monday. Can you make me an agent?"

The chat calls `validate_agent`. DutyGraph returns each scope as Inherited, Auto-trimmed or Escalated, with its rule. A scope no agent may hold is refused in the chat. If you continue, the chat calls `create_agent_request`, and the request goes to the approval chain in DutyGraph.

## What a request is not

A request is not an approval. An approval is not provisioning: your IAM team creates the agent's identity and access in your own systems. See [Approval, provisioning and revalidation](approval-provisioning-and-revalidation.md).

If no Task Card fits the work yet, a short capture of that one task, confirmed by your manager, is **Planned**. See [Just-in-time task capture](../roadmap/just-in-time-task-capture.md).

::: info
**Status.** **Built**: the four tools, the request card and token sign-in. **Partial**: task-based justification of each scope. **Planned**: OAuth sign-in, which web and mobile AI connectors need, CEIL-02 enforcement and just-in-time task capture.
:::
