Google Cloud MCP Toolbox Demo

This repository contains a small demo agent that uses Google Gemini embeddings and an MCP Toolbox MongoDB toolset to implement an "Online Groceries Agent". The agent demonstrates how to search for products in an inventory using vector search, add products to a cart (MongoDB collection) and get the cart total (aggregating the cart collection).

• mongodb-groceries-agent/: Contains the demo agent implementation (agent.py) and package init.
• tools.yaml: Toolbox toolset definition for the MongoDB tools.

Prerequisites

Before running the demo, ensure you have the following installed and available on your machine:

• Python 3.10+ (3.11 recommended)
• pip (or an equivalent Python package manager)
• MCP Toolbox Server (follow the instructions)
• A Google API key with access to the GenAI services (set as GOOGLE_API_KEY environment variable)
• A MongoDB deployment (Atlas free M0 cluster is supported — instructions below)

Setting up a free MongoDB Atlas M0 cluster

You can create a free (M0) MongoDB Atlas cluster. Follow these steps:

1. Open the Atlas registration page in your browser.

2. Sign up for a free MongoDB account or sign in if you already have one.

3. Click "Create a Cluster" and choose the free tier (M0 / Shared).

4. Choose a cloud provider and region (defaults are fine for the demo).

5. Create a database user: go to "Database Access" and add a user with a password. Save this username/password — you'll need it for connection strings.

6. Allow your IP address: go to "Network Access" and add your current IP or 0.0.0.0/0 for testing (not recommended for production).

7. Click "Connect", choose "Connect your application", and copy the provided connection string.

   Use the connection string in tools.yaml as the uri or preferrably, set it through an environment variables. An example connection string looks like:

   mongodb+srv://<username>:<password>@<your-cluster>.mongodb.net/<dbname>?retryWrites=true&w=majority
   

   Note: Replace <username>, <password>, <your-cluster> and <dbname> with values from Atlas. For local testing you can also use a local MongoDB server.

8. Import the sample grocery products data into a collection named products in a database named groceries. You can use the provided sample_data/groceries.json file and the MongoDB Atlas UI or mongosh to import it.

mongoimport --uri="<your-connection-string>" --db="grocery_store" --collection="inventory" --file="grocery_store.inventory.json" --jsonArray

Configuration

Set the following environment variables before running the agent:

• GOOGLE_API_KEY — your Google GenAI API key

Example (macOS / zsh):

export GOOGLE_API_KEY="your-google-api-key"

Running the demo agent

1. Install dependencies (adjust if this repo is part of a larger project). From the project root:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt  # if a requirements file exists; otherwise install needed packages manually

2. Start the toolbox server:

./toolbox --tools-file "tools.yaml"

3. Ensure the Toolbox server is running and accessible at http://127.0.0.1:5000.

4. Run the agent code (this demo file is intended to be imported by an orchestration script that starts the agent). To quickly try a simple run or import the module:

adk web

Adjust calls as needed to integrate the agent into a runtime environment that manages the Agent lifecycle.

What agent.py does

• Reads GOOGLE_API_KEY from the environment and sets it on os.environ for the Google GenAI client.
• Creates a genai.Client() and a Toolbox client targeting http://127.0.0.1:5000.
• Provides generate_embeddings(query) to produce embedding vectors using gemini-embedding-001.
• Provides find_similar_products(query) which runs a vector search via the toolbox tool find_similar_documents against an index named vector_index and the field gemini_embedding.
• Builds an Agent configured to use those tools and a grocery-shopping-toolset loaded from the toolbox.

Troubleshooting

• If you get authentication errors from GenAI, confirm GOOGLE_API_KEY is set and valid.
• If the Toolbox client cannot connect, confirm the ToolboxSync server is running and reachable at the configured address.
• If MongoDB connection issues appear, verify your Atlas IP allowlist and database user credentials.

Contributors ✨

Stanimira Vlaeva

Disclaimer

Use at your own risk; not a supported MongoDB product