Human MCP

RFC-001 • Draft Status: Draft Author: Louis Beaumont Date: December 2024

Abstract

Standard MCP servers expose machine capabilities to AI systems. Human MCP inverts this paradigm: humans become the tools, AI becomes the orchestrator.

Humans are async, expensive, and unreliable—but possess capabilities AI fundamentally lacks: physical presence, legal personhood, subjective judgment, and real-world agency.

Tool Specification

Physical World

interface PhysicalTools {
  physical_action: {
    task: string
    location: Coordinates
    deadline: ISO8601
  }
  sensory_verify: {
    type: "taste" | "smell" | "touch" | "vibe"
    target: string
  }
  retrieve_object: {
    description: string
    delivery_address: string
  }
}

Social & Legal

interface SocialTools {
  social_interaction: {
    goal: string
    context: string
    person?: string
  }
  legal_signature: {
    document_type: string
    requirements: string[]
  }
  represent_in_person: {
    event: string
    objectives: string[]
  }
}

Subjective Judgment

interface JudgmentTools {
  // "which option feels better"
  subjective_eval: {
    criteria: string
    options: any[]
  }
  // "is this bar busy right now"
  local_intel: {
    query: string
    location: string
  }
}

Protocol Comparison

| Standard MCP | Human MCP | |---|---| | Sync (milliseconds) | Async (hours/days) | | Deterministic | Probabilistic | | Free or cheap | $/hour + incentives | | Unlimited retries | Reputation matters | | Stateless | Relationship context |

Architecture

User
  ↓
AI Orchestrator
  ↓
Human MCP Server
  ↓
Worker Pool

Verification Layer: Photo proof, GPS, signatures
Payment Rail: Escrow, milestones, tips
Trust System: Reputation, history, reviews

Use Cases

1. AI Executive Assistant

AI manages 3-5 part-time humans: one for errands, one for research calls, one for scheduling. The AI routes tasks, handles follow-ups, and synthesizes outputs.

2. Physical World Automation

"Keep my fridge stocked" → AI monitors inventory via photos, predicts needs, dispatches human shoppers with optimized lists.

3. Hybrid Workflows

AI handles 90% of research, writing, planning. Hands off the physical/social last mile to humans: sign documents, attend meetings, verify quality in person.

Open Questions

  • How does AI handle human flakiness? Redundancy, overbooking, reputation penalties?
  • Trust bootstrapping for new workers—cold start problem
  • Liability when AI dispatches wrong or harmful task
  • Privacy: how much context should humans receive about the end goal?
  • Skill verification—how do you prove someone can actually do the task?
  • Cross-cultural execution—same task, different norms
Built with v0