Ask for Input

An MCP tool can stop and ask the user before it runs. Use that when the tool deletes a file or sends data.

Call ctx.context.requestInput inside the tool. The host shows your message to the user, and the call returns the answer.

Quick setup#

You need an MCP server first. If you do not have one, start with Serve Tools over HTTP.

1. Give .server() the type MCPToolContext, so ctx.context.requestInput type-checks.
2. Call requestInput with the message for the user.
3. Act on the answer.

import { toolDefinition } from '@tanstack/ai'
import { createMCPServer } from '@tanstack/ai-mcp/server'
import type { MCPToolContext } from '@tanstack/ai-mcp/server'
import { z } from 'zod'

const deleteFile = toolDefinition({
  name: 'delete_file',
  description: 'Delete a file after the user says yes',
  inputSchema: z.object({
    path: z.string(),
  }),
}).server<MCPToolContext>(async ({ path }, ctx) => {
  const answer = await ctx.context.requestInput({
    message: `Delete ${path}? Type yes to continue.`,
  })

  if (answer !== 'yes') {
    return `Kept ${path}.`
  }

  return `Deleted ${path}.`
})

const server = createMCPServer({
  name: 'files',
  version: '1.0.0',
  tools: [deleteFile],
})

export function fetch(request: Request) {
  return server.fetch(request)
}

If the user sends yes, the tool returns Deleted plus the file path. Any other answer returns Kept plus the path.

How it works#

• ctx is the second argument of the function that you pass to .server().
• requestInput returns the string that the user sends.
• If the user declines or cancels, requestInput throws an Error. The tool call then ends with a tool error.

A TanStack chat() host pauses on the request and shows it to the user. MCP Client Input shows how the user answers.

Advanced#

Spec 2025 and spec 2026#

The two MCP spec versions ask for input in different ways.

On spec 2025:

• requestInput needs sessions: 'memory'. Without a session, it throws. See MCP Server Sessions.
• requestInput waits on the open session. The same tool call then continues with the answer.

On spec 2026:

• The handler returns input_required. Then the client runs the tool again with the answer.
• The code before requestInput runs on each call, so it can run more than once. If work must run once, put it after requestInput returns.
• A tool asks one question per call. A second requestInput in the same call throws an Error. To ask two questions, use two tools.

Task tools#

requestInput does not work in an execution: 'task' tool. MCP Server Tasks has the details.