When Your AI Hits a Wall, Ask a Human

Your AI assistant is stuck. It needs a creative product name, subjective feedback on a design, or help with a complex decision that requires human intuition. Instead of throwing generic responses or admitting defeat, what if it could just... ask a person?

MCP-Human connects your AI assistant directly to Amazon Mechanical Turk workers, creating a seamless human-in-the-loop system through the Model Context Protocol. When your AI encounters questions it can't answer well, it posts them to real humans and waits for responses.

The Problem with Pure AI Workflows

AI assistants excel at many tasks but consistently struggle with:

• Creative brainstorming that needs fresh perspectives
• Subjective judgments about design, messaging, or user experience
• Complex decision-making that requires human context
• Domain-specific knowledge not in training data
• Tasks requiring emotional intelligence or cultural sensitivity

You could manually switch between AI and human help, but that breaks your workflow and requires constant context switching.

Real Human Intelligence, On Demand

This MCP server lets your AI assistant seamlessly escalate to human workers when needed. Set up once, then your AI can:

// Your AI assistant can now do this automatically
const response = await call("askHuman", {
  question: "What's a catchy name for a productivity app that helps remote teams stay focused?",
  reward: "0.25",
  title: "Creative product naming",
  hitValiditySeconds: 3600
});

Within minutes (or seconds), a real person reviews the question and provides their answer. Your AI gets the response and continues the conversation with human-backed insights.

Where This Actually Shines

Product Development: Get instant feedback on feature ideas, naming, or positioning from real users rather than AI guesses about market fit.

Content Creation: When your AI-generated copy feels generic, escalate to humans for authentic voice, cultural context, or creative alternatives.

Design Decisions: "Which of these three layouts feels more trustworthy for a financial app?" - questions that need human perception.

Research & Analysis: Gather qualitative insights on topics where human experience matters more than data synthesis.

Quality Control: Have humans verify AI-generated content for accuracy, tone, or appropriateness before publishing.

Simple Integration

Add it to Claude Desktop in one command:

claude mcp add human -- npx -y mcp-human@latest

The server defaults to MTurk's sandbox mode for testing. When you're ready for real responses, switch to production mode.

Your AI assistant immediately gains two new capabilities:

• askHuman - Send questions to MTurk workers with custom rewards and timeframes
• checkHITStatus - Poll for responses on longer-running tasks

Production-Ready Features

Cost Control: Set exact reward amounts per question (default $0.05, adjust based on complexity)

Time Management: Configure how long workers have to respond (1 hour default)

Sandbox Testing: Test your integration without spending real money using MTurk's sandbox

Status Tracking: Monitor active HITs and retrieve responses when ready

AWS Integration: Uses your existing AWS credentials and MTurk account setup

Getting Started

You'll need AWS credentials with MTurk permissions. The README includes step-by-step AWS setup instructions, but if you're already using AWS services, it's straightforward.

Configure your environment:

export AWS_ACCESS_KEY_ID="your_access_key" 
export AWS_SECRET_ACCESS_KEY="your_secret_key"

Then add the MCP server to your client. Start in sandbox mode to test, then switch to production when ready.

The Bigger Picture

This isn't just about getting human answers - it's about creating AI workflows that know their limitations and automatically seek help when needed. Instead of producing mediocre AI-only responses, your assistant can deliver genuinely helpful results by combining AI efficiency with human insight.

Whether you're building products, creating content, or solving complex problems, some questions genuinely need human perspective. Now your AI can get those answers without breaking your workflow.

Repository: github.com/olalonde/mcp-human