Skip to main content

​
connector.json vs MCP

​
Different Layers, Same Ecosystem

connector.json:  AI ↔ AI     (orchestration — who does what, when, where)
MCP:             AI → tool   (access — connect to a service, call a function)
They solve different problems and work together.

​
MCP: Tool Access Protocol

MCP (Model Context Protocol) lets AI connect to external tools and data: 
{
  "mcpServers": {
    "github": { "command": "npx", "args": ["@anthropic/mcp-github"] },
    "postgres": { "command": "npx", "args": ["@anthropic/mcp-postgres"] }
  }
}
  • AI calls github.create_pr() → MCP sends request → GitHub API → response
  • AI calls postgres.query() → MCP sends request → database → results
The target is passive. Tools don’t initiate. They respond.

​
connector.json: AI Orchestration Spec

connector.json lets AI coordinate with other AI: 
{
  "models": { "primary": "claude", "review": "codex" },
  "pipeline": {
    "phases": [
      { "name": "implement", "model": "primary" },
      { "name": "review", "model": "review" }
    ]
  }
}
  • Claude implements → Codex reviews → results synthesized
  • Both are active participants. Both reason, decide, act.
The target is active. Other AI agents that think and respond intelligently.

​
How They Compose

connector.json includes MCP as a tool configuration: 
{
  "connector": "2.0",
  "models": { "primary": "claude" },
  "tools": {
    "mcp": ["github", "postgres", "chrome-devtools"]
  },
  "pipeline": {
    "phases": [
      {
        "name": "investigate",
        "model": "primary",
        "prompt": "Use GitHub MCP to find recent PRs. Use Postgres MCP to check migration status."
      }
    ]
  }
}
MCP servers are available inside connector.json sessions. The AI worker can use MCP tools while participating in the connector pipeline.

​
Comparison Table

MCPconnector.json
What it connectsAI → tools/dataAI ↔ AI
TargetPassive (responds)Active (reasons)
DirectionOne-way request-responseMulti-directional collaboration
ProtocolJSON-RPC over stdio/SSEJSON spec → PTY sessions
Session isolationNot applicableworktree, sandbox per session
Cost routingNot applicableRoute by model cost/capability
FailoverNot applicableSwitch to different LLM
Multi-phaseNot applicablePipeline with sequential phases
Cross-machineServer can be remoteWorkers on different machines
Standard bodyAnthropic-ledOpenCLIs / Monolex

​
Layer Diagram

┌─────────────────────────────────────────────┐
│  connector.json (orchestration)             │
│  Models, pipeline, session isolation        │
│  ┌─────────────────────────────────────┐    │
│  │  MCP (tool access)                  │    │
│  │  github, postgres, chrome, slack    │    │
│  └─────────────────────────────────────┘    │
│  ┌─────────────────────────────────────┐    │
│  │  PTY-for-AI (execution)             │    │
│  │  daemon, headless, session plugins  │    │
│  └─────────────────────────────────────┘    │
└─────────────────────────────────────────────┘
connector.json sits on top. MCP and PTY-for-AI are execution layers inside it.