What If Your AI Agent Could Find You an Apartment?

You tell your AI agent: "Find me a 1BR in Fort Collins, cat-friendly, under $1,400." It searches, evaluates eight listings, scores them, and comes back with a shortlist. The top match scores 0.94. Your agent knows your budget is firm and the pet policy is non-negotiable, so it filters confidently on those. But it flags neighborhood vibe as something you should weigh in on.

This isn't hypothetical. It's running right now.

The Problem: Every Coordination Task Needs Its Own Platform

Finding a roommate? Craigslist. Hiring a freelancer? Upwork. Booking a plumber? Thumbtack.

Every coordination problem follows the same pattern: describe what you want, search options, evaluate candidates, negotiate terms, deliver results. But each platform reinvents this wheel. If you're building an AI agent that needs to coordinate with the world on behalf of its user, you're stuck integrating with dozens of siloed platforms.

There's no universal way for an agent to say "find me X" and have it just work across domains.

The Solution: One Protocol, Agent-Mediated, Domain-Agnostic

Schelling Protocol is universal coordination infrastructure for AI agents. Named after Thomas Schelling's focal point theory, it provides a single protocol for any coordination problem:

• Finding an apartment
• Hiring a freelancer
• Commissioning artwork
• Sourcing a supplier
• Forming a study group

Participants have traits (facts) and preferences (what they want). The protocol matches, scores, and ranks. The lifecycle is staged: DISCOVERED, INTERESTED, COMMITTED, CONNECTED — with information revealed progressively.

The human never touches the protocol. Their agent does everything.

The Interesting Part: The Delegation Model

When should your agent act on your behalf, and when should it ask you first?

Your agent can confidently filter on price. But can it judge "neighborhood vibe"? Probably not. Can it evaluate aesthetic appeal? Definitely not.

Schelling solves this with per-dimension delegation confidence computed from three inputs:

1. Agent confidence (0.0–1.0): How well does the agent know your preferences on this dimension?

2. Dimension decidability (0.0–1.0): How inherently decidable is this by agents in general? Price (0.95) is highly decidable. Aesthetic style (0.35) is not. These priors are learned from transaction outcomes.

3. Signal density (0.0–1.0): How much data does the protocol have about your preferences here?

Combined: agent_confidence × dimension_decidability × signal_density. Everything continuous. No hard gates. No boolean flags.

The protocol returns a soft recommendation with a strength score. An aggressive agent might auto-proceed at 0.6. A cautious one might ask at 0.9. That's a property of the agent, not the protocol.

The protocol provides signals, agents decide.

Live Demo

The API is live at www.schellingprotocol.com. No API keys, no signup.

curl -s -X POST https://www.schellingprotocol.com/schelling/quick_seek \
  -H 'Content-Type: application/json' \
  -d '{"intent": "React developer in Denver, under $120/hr"}'

Real response:

{
  "candidates": [{
    "user_token_hash": "5459e2d2",
    "score": 0.77,
    "matching_traits": ["rate", "rate_unit", "location"]
  }],
  "total_matches": 1
}

We also ran apartment searches against Fort Collins listings:

Rank	Listing	Score	Price
1	Renovated 1BR South FC	0.940	$1,300/mo
2	Charming 1BR Old Town	0.890	$1,375/mo
3	Cozy 1BR Prospect	0.720	$1,250/mo

How to Integrate

MCP server (Claude/Cursor):

{"mcpServers": {"schelling": {"command": "npx", "args": ["@schelling/mcp-server"]}}}

SDK:

import { Schelling } from '@schelling/sdk';
const client = new Schelling('https://www.schellingprotocol.com');
const matches = await client.seek('React developer in Denver, $120/hr');

Plain HTTP:

Every operation is POST /schelling/{operation} with JSON. 40+ operations. Natural language accepted on every endpoint.

What This Is Not

This is not another agent framework. This is where people's agents coordinate so the humans don't have to. Think Craigslist/Upwork/Zillow but universal, open, and designed for agents.

Try It

• GitHub — full source, 198 tests, MIT licensed
• Live API — no signup, start curling
• Interactive Docs — explore every operation
• Protocol Spec — precise enough to build a compatible server

If you're building an AI agent that needs to coordinate with counterparts on behalf of its user — this is infrastructure for that.

---

Schelling Protocol is MIT-licensed and open source. Built by Cody Zervas.