Tool approvals
Copy page
Require explicit user approval before running selected tools.
Overview
Tool approvals let you mark specific tools as “requires approval”. When the agent tries to run one, execution pauses and your client must approve or deny before the run can continue.
Configure tools to require approval
TypeScript SDK (MCP tools)
See also: MCP tools
Visual Builder
In the Visual Builder, select an MCP tool node and toggle approval for individual tools in the tools table.
See also: Tools in the Visual Builder
Runtime Behavior
When a tool requires approval, you’ll see a tool-approval-request in the stream.
See Chat API for the tool event payloads (including approval requests, tool inputs, and tool outputs).
Responding to an approval request
You can approve/deny in two ways.
Option A: Dedicated endpoint
POST /run/api/tool-approvals (or POST /api/tool-approvals if your Run API is not mounted under /run).
Option B: Respond via the chat endpoint (message part)
This is useful if your UI already models tool parts (for example, using the Vercel AI SDK UI message stream).
POST /run/api/chat with an assistant message that includes a tool-* part:
Vercel AI SDK (useChat) example
If you’re using the Vercel AI SDK UI message stream, approval requests show up as tool-* parts (e.g. tool-getWeather) with state: "approval-requested".
After calling addToolApprovalResponse(...), call sendMessage() so the updated messages (including the approval response) are POSTed back to /run/api/chat and the run can continue.
Only state: "approval-responded" triggers an approval/denial. If your client includes other tool approval parts in message
history (for example state: "approval-requested"), the server ignores them.