Python: [Bug]: Support MCP tool binding by reference in declarative YAML #4927

Description

MAF declarative workflows require embedding full MCP connection details directly in the YAML tool definition. Two problems make this unworkable:

1. Custom HTTP headers are not supported in the YAML schema - only url is available. (Tracked in Python: [Bug]: Support custom HTTP headers in declarative YAML MCP tool definitions #4582)
2. There is no way to reference an MCP connection by identifier - connection details must be redeclared in every agent and workflow that uses the same MCP server, coupling sensitive infrastructure config to business workflow files.

The expected model is that MCP connections are declared once in configuration or code, and workflow YAML references them by name, declaring only which tools each agent is allowed to use.

Current Behaviour

Passing a pre-built MCP tool via bindings is ignored. MAF constructs the tool directly from the YAML spec regardless. Results in server_url: '' being sent to the API, returning a 500.

Note: McpTool does accept a bindings field (inherited from Tool) but _parse_tool in AgentFactory never reads it for the MCP case, unlike FunctionTool where bindings correctly resolve the callable.

The only working workaround is passing the tool at agent.run() with no tools block in the YAML, but this loses per-agent tool scoping entirely.

Code Sample

# Minimal agent YAML — no url, no headers, no connection details.
# Just the name and the allowed tools.
AGENT_YAML = """
kind: Prompt
name: TestAgent
instructions: You are a helpful assistant.
model:
    id: gpt-4.1
    provider: AzureOpenAI.Responses
tools:
    - kind: mcp
      name: product-mcp
      allowedTools:
        - customer_search
      # connection resolved from bindings by name at runtime
"""

async def main():
    mcp_tool = OpenAIResponsesClient.get_mcp_tool(
        name="product-mcp",
        description="Product MCP server",
        url="http://example.com",
        headers={"X-API-Key": "..."},
        approval_mode="never_require",
        allowed_tools=["customer_search"],
    )

    factory = AgentFactory(
        safe_mode=False,
        bindings={"product-mcp": mcp_tool},
    )

    async with factory.create_agent_from_yaml(AGENT_YAML) as agent:
        response = await agent.run("Search for customer Inforshow")
        print(f"Response: {response.text}")

Error Messages / Stack Traces

Package Versions

agent-framework: 1.0.0rc4, agent-framework-declarative: 1.0.0b260311

Python Version

3.13

Additional Context

From our understanding, Azure Foundry solves this cleanly via project_connection_id: MCP connections are registered centrally in the project (including headers and secrets) and referenced by name in the YAML:

tools:
  - type: mcp
    server_label: ERP_test_MCP
    server_url: http://example.com/
    require_approval:
      never:
        tool_names:
          - customer_search
    project_connection_id: ERP-test-MCP

Could a similar behavior be achievable through MAF, with centralized code definitions?