# Ask an AI assistant

Everything in the UI is also available to Claude, ChatGPT, GitHub Copilot, Cursor and any
other client that speaks the Model Context Protocol (MCP). Five minutes to the first answer.

## 1. Mint a personal access token

**Profile → API keys → New key.** Copy the `dq_live_…` token now; it is shown once. The token
acts as you — the same workspace role and the same per-suite `view` / `edit` grants — so an
assistant can never see or change more than you can.

![The Profile page, where personal access tokens are minted and revoked](../assets/screenshots/profile.png){ .screenshot }

*Tokens are listed by prefix only; revoke one here and it stops working on its next call.*

## 2. Point your client at `/mcp`

The server lives at `https://<your-dataq-host>/mcp/` (trailing slash) and authenticates with
the token as a bearer. Copy-paste configuration for Claude Desktop, Claude.ai, VS Code /
Copilot and Cursor is in [AI assistants (MCP)](../guides/mcp-setup.md).

## 3. Ask

Try the four questions the tools were designed around:

- *"Which suites failed in the last 24 hours, and why?"*
- *"Is the `orders` table healthy? What feeds it?"*
- *"Add a not-null check on `customer_id` to the orders suite and dry-run it."*
- *"Why did I not get an alert for last night's failure?"*

Every tool states what it **cannot** see — a mid-run suite is reported as *not final*, an
expired snooze as *expired*, a truncated list as *truncated* — because an assistant has no
"running" badge to glance at. That design is documented in
[MCP tool design](../architecture/mcp-honesty.md), and the full list of the 48 tools with who
can call them is the [MCP tools reference](../reference/mcp-tools.md).

!!! tip "Ask about the docs, too"
    Every page on this site has **Ask AI** at the top: it opens the page in your own assistant
    with the raw Markdown, and the same pages are available to assistants through the `get_doc`
    tool.