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