Lab 27: Using a Stateful File System Tool Challenge

Goal

In this lab, you will learn how to connect your ADK agent to an external, stateful tool using the Model Context Protocol (MCP). You will use the McpToolset to connect to a pre-built, open-source MCP server that provides file system operations. This will allow your agent to list files and read their contents from your local machine.

Prerequisites

Node.js and npx: The MCP server we will use is a Node.js package. If you don't have it, install Node.js (which includes npx) from the official website.

Step 1: Create the Agent Project

Navigate to your training directory:
```
cd /path/to/your/adk-training
```
Create the agent project:
```
uv run adk create mcp_agent
cd mcp_agent
```
Install the MCP extra. mcp is not part of ADK's base install — it only comes in via ADK's [mcp] extra, which pins a compatible version:
```
uv add "google-adk[mcp]"
```

Step 2: Create a Test Directory and File

The MCP file system server needs a directory to operate on. Let's create one.

Create a directory for the tool to access: Inside your mcp_agent project, create a directory named test_files.
```
mkdir test_files
```
Create a sample file: Inside test_files, create a file named hello.txt.
```
echo "Hello from the MCP world!" > test_files/hello.txt
```

Step 3: Configure the Agent to Use the MCP Toolset

Because the McpToolset requires Python code to configure the connection, we must define our agent in an agent.py file.

Create the agent.py file: In the mcp_agent directory, create a file named agent.py.
Create the __init__.py file: This empty file is crucial. It tells Python to treat the mcp_agent directory as a package, allowing uv run adk web to discover and load your agent.py.
```
touch __init__.py
```

Complete the agent.py script: Exercise: Open agent.py and complete the script by following the # TODO comments. Your goal is to define an agent and configure the McpToolset to launch and connect to the file system server.

import os
# TODO: 1. Import the necessary classes:
from google.adk import Agent
from google.adk.tools.mcp_tool.mcp_toolset import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StdioConnectionParams
from mcp import StdioServerParameters

# -- Configuration --
# TODO: 2. Define the TARGET_FOLDER_PATH.
TARGET_FOLDER_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "test_files/")

# -- Agent Definition --
# TODO: 3. Define the root_agent.
# Instantiate an Agent with the following properties:
# - model: 'gemini-3.5-flash'
# - name: 'filesystem_agent'
# - instruction: 'You are a helpful assistant that can interact with a user\'s local file system. You can list files and read their content.'
# - tools: A list containing one item: the McpToolset.

# Inside the Agent's `tools` list, you will configure the McpToolset.
# Follow this structure:
#
# McpToolset(
#     connection_params=StdioConnectionParams(
#         server_params=StdioServerParameters(
#             # TODO: 4. Set the `command` to 'npx'.
#             # TODO: 5. Set the `args` to a list containing:
#             # "-y", "@modelcontextprotocol/server-filesystem", absolute path to TARGET_FOLDER_PATH
#             command=...,
#             args=[...],
#         ),
#     ),
#     # TODO: 6. Optionally, filter specific tools: 'list_directory' and 'read_file'.
#     tool_filter=[...]
# )

root_agent = ...

Set up your .env file with your API key or Agent Platform project.

Step 4: Test the Stateful Tool

Navigate to the parent directory (cd ..) and start the Dev UI:
```
uv run adk web
```
When the server starts, you will see output in the console as npx downloads and runs the @modelcontextprotocol/server-filesystem package.

Note: if your very first request fails with an MCP session timeout, this is likely npx downloading the server package for the first time. Simply retry — it will be fast on every subsequent run since the package stays cached locally.
Interact with the agent:
- Open the Dev UI in your browser.
- Select the filesystem_agent from the dropdown.
- Turn 1: List the files.
  - User: "What files are in my directory?"
  - Expected Response: The agent should respond with a message indicating that it sees hello.txt.
- Turn 2: Read the file.
  - User: "Great, can you read the content of hello.txt for me?"
  - Expected Response: The agent should respond with the content of the file: "Hello from the MCP world!"
Examine the Trace View:
- In the trace for both turns, you will see execute_tool steps for list_directory and read_file. These tools were dynamically discovered by the McpToolset.

Having Trouble?

If you get stuck, you can find the complete, working code in the lab-solution.md file.

Lab Summary

You have successfully connected your ADK agent to a stateful, external tool using the Model Context Protocol.

You have learned to:

Understand the client-server architecture of MCP.
Use the McpToolset to connect to an MCP server.
Configure the StdioConnectionParams to automatically launch a local MCP server process directly within your agent definition.
Build an agent that can use tools provided by an external service without having to define them locally.

Bonus: Connecting to a Remote MCP Server

So far you connected to a local MCP server launched as a subprocess (StdioConnectionParams). Most real-world integrations instead connect to a server already running somewhere else over HTTP, using StreamableHTTPConnectionParams — same McpToolset, different transport.

Get a free GitHub Personal Access Token: create one at github.com/settings/personal-access-tokens/new with read-only repository access.
Set it as an environment variable in your .env file:
```
GITHUB_TOKEN=YOUR_GITHUB_TOKEN
```

Create a second agent (e.g. github_agent/agent.py):

import os
from google.adk import Agent
from google.adk.tools.mcp_tool import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StreamableHTTPConnectionParams

GITHUB_TOKEN = os.environ["GITHUB_TOKEN"]

root_agent = Agent(
    model="gemini-3.5-flash",
    name="github_agent",
    instruction="Help users get information from GitHub repositories.",
    tools=[
        McpToolset(
            connection_params=StreamableHTTPConnectionParams(
                url="https://api.githubcopilot.com/mcp/",
                headers={
                    "Authorization": f"Bearer {GITHUB_TOKEN}",
                    "X-MCP-Readonly": "true",
                },
            ),
        )
    ],
)

Try it: ask "What are the open issues on google/adk-python?" — no npx, no subprocess, no local sandboxing: the tool call goes straight over HTTPS to GitHub's servers.

A few things change when the server is remote instead of local: there's no subprocess lifecycle to manage (the server runs independently of your agent), network failures (timeouts, rate limits, an expired token) become a real possibility that a local stdio server never has, and the security surface shifts from "the subprocess has filesystem access" to "don't hardcode the credential in your header" — which is why the token above comes from .env, not a literal string.

Self-Reflection Questions

The McpToolset dynamically discovers the tools from the server. What are the advantages of this approach compared to manually defining each tool on the agent side?
The file system server is "stateful" because it remembers the state of the test_files directory between tool calls. How does this differ from the stateless calculator tools you built in earlier modules?
The StdioConnectionParams launches the MCP server as a subprocess. What are the security implications of this, and why is it important that the server is sandboxed to a specific TARGET_FOLDER_PATH?

🕵️ Hidden Solution 🕵️

Looking for the solution? Here's a hint (Base64 decode me): L2RvYy1hZGstdHJhaW5pbmcvbW9kdWxlMjctaW50cm8tdG8tbWNwL2xhYi1zb2x1dGlvbg==

The direct link is: Lab Solution

Goal​

Prerequisites​

Step 1: Create the Agent Project​

Step 2: Create a Test Directory and File​

Step 3: Configure the Agent to Use the MCP Toolset​

Step 4: Test the Stateful Tool​

Having Trouble?​

Lab Summary​

Bonus: Connecting to a Remote MCP Server​

Self-Reflection Questions​

🕵️ Hidden Solution 🕵️​