> Call the `diagnose` tool first. Full index: /llms.txt. Errors: /reference/problems.

# Answer cards

A [card](/concepts/cards) waits for an answer. Answer it from the web app or through your agent.

## In the web app

1. Open **Inbox**. A number on it counts what needs you.
2. Open a card. Read the question and what depends on it.
3. Answer in your own words, provide the value, or skip the card.
4. Check the card's outcome. It shows whether the answer was applied.

## Through your agent

```text
requests(action="list", project_id)
requests(action="list", project_id, since_cursor=<next_since_cursor>)
requests(action="answer_card", project_id, ...)
```

The second call returns only the cards opened or settled since the last one, so poll it between turns. `watch_run` also reports `card_opened` and `blocked` events as they happen.

Your agent may answer a card about a missing login or value with data you gave it. A card about what the app should do gets only your own words. If you did not say it, the agent asks you.

## After the answer

An answered card is not an applied one. Read its outcome first. Then start the next run yourself: nothing regenerates or runs on its own.

See [MCP tools](/reference/mcp/requests-dev) for the arguments.