MCP Tool Response Role Incorrectly Set as "user"

[pansusu]

Issue: MCP Tool Response Role Incorrectly Set as "user"

Description

When using ADK to call MCP server through MCPToolset, the tool response's role is incorrectly set as "user" instead of "assistant". This appears to be inconsistent with the expected behavior of tool responses.

Current Behavior

# Example response format
{
    "parts": [
        {
            "function_response": {
                "id": "call_xxx",
                "name": "read_query",
                "response": {
                    "result": {
                        "content": [...],
                        "isError": false
                    }
                }
            }
        }
    ],
    "role": "user"  # <-- This should be "assistant"
}

Expected Behavior

Tool responses should have "role": "assistant" to properly indicate that they are responses from the agent/tool rather than user input.

Environment

• ADK version: [version]
• MCP server: SQLite implementation
• Python version: [version]

Steps to Reproduce

1. Initialize MCPToolset with SQLite server
2. Make a tool call through the agent
3. Observe the response format

Impact

This issue might affect:

• Message flow tracking
• Response handling logic
• UI rendering of conversation history

Additional Context

The issue occurs when using MCPToolset.from_server() to connect to the MCP server. The response format suggests there might be a mismatch between MCP server's response format and ADK's expected format.

Questions

1. Is this the intended behavior of MCPToolset?
2. Are there configuration options to control the response role?
3. Should this be handled at the MCP server level or ADK level?

Related Code

local_tools, _ = await MCPToolset.from_server(
    connection_params=StdioServerParameters(
        command='uv',
        args=[...],
    ),
    async_exit_stack=self._exit_stack
)

[calvingiles]

Tool calls should be set to assistant / model, but tool responses are provided from outside the model, and are considered a form of user input by at least several providers, including gemini and claude - so I think the current behaviour is the intended behaviour.

[seanzhou1023]

@calvingiles is correct.

This is by design. Tool call is issued by model, while tool execution is not executed by model and thus tool response has role as 'user'