Google Adk MCP Tool

Google ADK — MCP Tool Integration

Import

from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset
from mcp import StdioServerParameters

Install MCP Extension

pip install google-adk[mcp]
# or
pip install google-adk[extensions]

Stdio Transport (Local MCP Server)

from google.adk.agents import Agent
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset
from mcp import StdioServerParameters

notion_toolset = MCPToolset(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@notionhq/notion-mcp-server"],
        env={"NOTION_API_KEY": "your-key-here"},
    ),
)

root_agent = Agent(
    name="notion_agent",
    model="gemini-2.5-flash",
    instruction="Help users manage their Notion workspace.",
    tools=[notion_toolset],
)

SSE Transport (Remote MCP Server)

from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset, SseServerParams

remote_toolset = MCPToolset(
    connection_params=SseServerParams(
        url="http://localhost:8080/sse",
        headers={"Authorization": "Bearer token"},
    ),
)

root_agent = Agent(
    name="remote_agent",
    model="gemini-2.5-flash",
    instruction="Use the remote tools to complete tasks.",
    tools=[remote_toolset],
)

Tool Filtering

Only expose specific tools from the MCP server:

toolset = MCPToolset(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@modelcontextprotocol/server-filesystem"],
    ),
    tool_filter=["read_file", "write_file", "list_directory"],
)

Tool Filter with Callable

def my_filter(tool) -> bool:
    """Only allow read operations."""
    return tool.name.startswith("read_") or tool.name.startswith("list_")

toolset = MCPToolset(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@modelcontextprotocol/server-filesystem"],
    ),
    tool_filter=my_filter,
)

Multiple MCP Servers

filesystem_tools = MCPToolset(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@modelcontextprotocol/server-filesystem"],
    ),
)

github_tools = MCPToolset(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@modelcontextprotocol/server-github"],
        env={"GITHUB_TOKEN": "your-token"},
    ),
)

root_agent = Agent(
    name="dev_agent",
    model="gemini-2.5-flash",
    instruction="Help with development tasks using filesystem and GitHub.",
    tools=[filesystem_tools, github_tools],
)

MCP with Agent Instructions

Use McpInstructionProvider to fetch agent instructions from an MCP server's prompts:

from google.adk.agents import Agent
from google.adk.agents.mcp_instruction_provider import McpInstructionProvider
from mcp import StdioServerParameters

# Fetches a named prompt from the MCP server as the agent instruction
mcp_instructions = McpInstructionProvider(
    connection_params=StdioServerParameters(
        command="npx",
        args=["-y", "@my-org/my-mcp-server"],
    ),
    prompt_name="agent_instructions",
)

root_agent = Agent(
    name="my_agent",
    model="gemini-2.5-flash",
    instruction=mcp_instructions,  # Callable that returns str
    tools=[notion_toolset],
)

Common MCP Servers

Key Rules

• MCP tools require pip install google-adk[mcp]
• Stdio transport starts a local process — ensure the binary/npx package exists
• SSE transport connects to a running remote server
• Tool filtering prevents agents from accessing dangerous tools
• MCPToolset is async — works with adk web, adk run, and Runner.run_async()
• Environment variables in StdioServerParameters.env are passed to the MCP process

Related Skills

• google-adk-function-tool — Custom function tools
• google-adk-openapi-tool — REST API tools from OpenAPI specs