Answers and approvals (“Needs you”)

What reaches you

• A question an agent cannot settle by itself, for example “which of these two ways?”
• A permission request: an agent wants to run something its AI program asks you about.
• A result waiting for your accept (see Items and the board).
• A request for something only the person can do: it arrives as one request with a button.

Where to look. Open Home, the page you land on. “Needs you” groups what waits: “Answer a question” (an item stuck on you), “Accept work”, “Deferred”, “Answer an agent” (an agent waits until you answer) and “For information”. A pop-up “Decision N of M” shows requests one at a time; “Later” snoozes one for an hour.

How to answer

1. Press the option you want. harnsy answers in the agent’s own session, and the agent goes on. The terminal keeps working; the first answer counts.
2. For a result, open the card and press “Accept” or “Return…”.

Where it works: for a Claude Code agent, both questions and permission requests; for Codex, questions only; for OpenCode, both. Claude Code permission requests need harnsy’s answer hook, which the installer sets up.

Talk to a live agent. For a Claude Code or Codex agent on your computer, open its chat: you see its tool calls, commands and edits as they happen. What only you can do appears in the chat as a request with a button.

How agents start: the shield. The shield in the top bar shows how agents are started. Amber: you have not chosen. Green: with permission checks. Red: without.

Tools you connected (MCP servers). You grant tools per project and role. Agents call the ticked tools without asking for each write. These calls wait for your “Yes, run it”: when “Confirm every write” is on; when the server is under strict control and the tool may delete data; when a tool is marked as deleting; when the arguments carry a secret of your machine; and after 30 writes through one connection within 5 minutes. An undecided call expires after 24 hours and the agent is told. See MCP servers.