MCP Integration¶

DeepFabric supports the Model Context Protocol (MCP) for tool definitions. Load tools from MCP servers or use MCP-format schemas directly.

What is MCP?¶

MCP is a standard for defining tools that AI models can use. It specifies: - Tool names and descriptions - Input schemas (JSON Schema format) - Annotations for tool behavior (read-only, destructive, etc.)

Loading from MCP Server¶

Pull tools directly from an MCP server:

curl -X POST http://localhost:3000/mock/pull \
  -H "Content-Type: application/json" \
  -d '{"url": "http://your-mcp-server:8000"}'

The Mock component sends a tools/list JSON-RPC request and loads the returned tools.

MCP Tool Format¶

{
  "name": "search_code",
  "description": "Search code in a repository",
  "inputSchema": {
    "type": "object",
    "properties": {
      "repository": {
        "type": "string",
        "description": "Repository name"
      },
      "query": {
        "type": "string",
        "description": "Search query"
      }
    },
    "required": ["repository", "query"]
  },
  "annotations": {
    "readOnlyHint": true
  }
}

Annotations¶

MCP annotations describe tool behavior:

Annotation	Description
`readOnlyHint`	Tool doesn't modify state
`destructiveHint`	Tool may be destructive
`idempotentHint`	Safe to retry
`openWorldHint`	Interacts with external systems

DeepFabric Tool Format¶

DeepFabric also uses its own format internally:

- name: search_code
  description: Search code in a repository
  parameters:
    - name: repository
      type: str
      description: Repository name
      required: true
    - name: query
      type: str
      description: Search query
      required: true
  returns: "List of matching files and line numbers"

Conversion¶

DeepFabric automatically converts between formats: - MCP schemas are converted for internal use - Output datasets use OpenAI tool format

Configuration¶

Using tools_endpoint¶

Load tools from an HTTP endpoint returning MCP format:

generation:
  tools:
    spin_endpoint: "http://localhost:3000"
    tools_endpoint: "http://localhost:3000/mock/list-tools"
    tool_execute_path: "/mock/execute"

Custom Tools in Config¶

Define tools directly in YAML:

generation:
  tools:
    custom:
      - name: send_email
        description: Send an email
        parameters:
          - name: to
            type: str
            required: true
          - name: subject
            type: str
            required: true
          - name: body
            type: str
            required: true
        returns: "Confirmation message"

Example Workflow¶

Start MCP server with your tool definitions

Load into Mock component:

curl -X POST http://localhost:3000/mock/pull \
  -d '{"url": "http://localhost:8000"}'

Configure DeepFabric:

tools:
  spin_endpoint: "http://localhost:3000"
  tools_endpoint: "http://localhost:3000/mock/list-tools"

Generate dataset with your MCP tools

Output Format¶

Generated datasets include OpenAI-format tool definitions:

{
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "search_code",
        "description": "Search code in a repository",
        "parameters": {
          "type": "object",
          "properties": {
            "repository": {"type": "string"},
            "query": {"type": "string"}
          },
          "required": ["repository", "query"]
        }
      }
    }
  ]
}

This format is compatible with OpenAI, TRL, and most training frameworks.