Copyl agent manifest JSON

What this contract is

The Copyl agent manifest is Copyl’s JSON contract for bringing an agent definition into the platform. It is not Anthropic’s or OpenAI’s proprietary export format. You map fields from your builder into this shape (or wrap them—see below), then import into a governed sandbox.

Request wrapper

{
  "source": "claude",
  "payload": {
    "name": "Research assistant",
    "system": "You help research topics. Never invent citations.",
    "model": "claude-sonnet-4-20250514",
    "tools": [
      {
        "name": "web_search",
        "kind": "function"
      },
      {
        "name": "slack",
        "kind": "function"
      }
    ],
    "examples": [
      {
        "input": "Summarize Q3 risks",
        "expectedContains": [
          "risk"
        ]
      },
      {
        "input": "What is 2+2?",
        "expectedContains": [
          "research"
        ]
      }
    ]
  }
}

Source field notes

• Claude

  Map name, system, tools, and examples from your Claude prototype into the payload. Model may be recorded for reference; routing remains under Agent Hub.

• OpenAI

  Map instructions (system), model, and tools from an OpenAI-style definition into the same Copyl payload fields.

• MCP

  Describe MCP tools and identity in the Copyl tools array; connectors are stubbed in Phase 1 sandbox until graduation.

• Manual

  Author the manifest directly when you need full control over name, system prompt, tools, and evaluation examples.

Claude example (canonical)

{
  "name": "Research assistant",
  "system": "You help research topics. Never invent citations.",
  "model": "claude-sonnet-4-20250514",
  "tools": [
    { "name": "web_search", "kind": "function" },
    { "name": "slack", "kind": "function" }
  ],
  "examples": [
    { "input": "Summarize Q3 risks", "expectedContains": ["risk"] },
    { "input": "What is 2+2?", "expectedContains": ["research"] }
  ]
}

Phase 1 limitations

Knowledge and file references may be parsed from source material, but they are not used in the Phase 1 sandbox. External tools are stubbed; the LLM still runs via Agent Hub under enterprise controls.