MCP-I

Introduction

— Servers prove their identity.
— Authorization enforced per-tool.
— Proof of every agentic action.

1. How It Works

When an AI agent wants to use a tool, both sides prove who they are—like showing ID before entering a building.

2. The Handshake

Before any tool is called, the server cryptographically proves its identity. No more trusting random endpoints.

{
  "capabilities": {
    "identity": {
      "challenge": "prove-yourself-123"
    }
  }
}

{
  "identity": {
    "did": "did:key:z6Mk...",
    "credential": "eyJ...",
    "proof": {
      "signature": "z58DAdFfa9..."
    }
  }
}

3. Tool Authorization

Tools declare what permissions they need.
The agent checks if the human granted those permissions before calling.

4. Every Action is Signed

When Claude connects, calls a tool, or returns a result, it includes  cryptographic proof.
An identity-signed receipt that can be audited later.

{
  "name": "read_calendar",
  "arguments": { "date": "2025-01-29" },
  "authorization": {
    "delegation": "eyJ...",
    "proof": {
      "timestamp": 1735689600,
      "signature": "z58DAdFfa9..."
    }
  }
}

5. Error Codes

-32001  UNAUTHORIZED          Missing/invalid credential
-32002  FORBIDDEN             Valid credential, insufficient scope
-32003  DELEGATION_EXPIRED    Delegation has expired
-32004  DELEGATION_REVOKED    Delegation was revoked

6. Implementation

Reference implementations:

npm install @kya-os/mcp-i-core          # Core runtime
npm install @kya-os/mcp-i-cloudflare    # Cloudflare Workers adapter
npx @kya-os/create-mcpi-app my-server   # Project scaffolder