Approvals

Human sign-off in front of sensitive work.

No tool creates an approval. One opens when a permission rule says ask (require_approval): the call stops with approval_required and an approval id, the approval carries the tool and its arguments, and it is assigned to the person who made the call — or, for an agent, to the agent's owner. Once they approve it, the same call goes through. Agents can read the approvals they asked for, but never approve or reject.

4 tools — 2 read, 2 write.

Tool		What it does
approve	write	Approve a pending approval.
get_approval	read	Get full details of an approval by ID, including status, assignee, resolution, and context.
list_approvals	read	List the approvals you requested or are assigned (for an agent: the ones it asked for), optionally filtered by status, assignee, or requester.
reject	write	Reject a pending approval.

Calling these

Over MCP, find one with search_tools, then run it with execute_tool. Over REST, POST /api/capabilities/<tool> with the arguments as the JSON body. Either way, fetch the input schema for the tool you're about to call rather than guessing — get_tool_schema over MCP, or GET /api/capabilities?name=<tool> over REST.

See How the API works for auth, errors and the shared envelope.