Skip to content

README.md

Latest commit

 

History

History
115 lines (83 loc) · 3.7 KB

README.md

File metadata and controls

115 lines (83 loc) · 3.7 KB

Using Smolagents with Mapbox MCP

This example demonstrates how to integrate Mapbox's Model Context Protocol (MCP) server with Smolagents, allowing AI agents to access Mapbox's location-based tools.

Overview

The smolagents_example.py script shows a simple but powerful implementation of connecting an AI agent to Mapbox's MCP server. It enables the agent to perform location-based tasks such as:

  • Getting directions between landmarks
  • Searching for points of interest
  • Geocoding locations
  • Calculating travel times and distances
  • Generating static map images
  • And more...

Prerequisites

  • Python with smolagents and mcp packages installed:
    • Option 1: pip install 'smolagents[mcp]'
    • Option 2: pip install -r requirements.txt (from this directory)
  • A Mapbox access token (set as an environment variable)
  • Node.js or docker (to run the MCP server)

With NPM package, you don't need to clone this repository, and build it. But If you want to use the local codes, you need to clone and build it:

# Build node (from repository root)
npm run build

# note your absolute path to node, you will need it for MCP config
# For Mac/Linux
which node
# For Windows
where node

# Alternatively, build docker
docker build -t mapbox-mcp-server .

How It Works

The script demonstrates different ways to configure language models:

  1. InferenceClient: For hosted models
  2. Transformers: For local models through HuggingFace
  3. Ollama: For local models through Ollama
  4. LiteLLM: For accessing various API-based models
  5. OpenAI: For OpenAI's models

It connects to the Mapbox MCP server, which exposes Mapbox's functionality as tools that the AI agent can use to answer location-based questions.

Important: The example uses structured_output=True when connecting to MCP, which enables smolagents to properly handle the structured data returned by Mapbox tools. This allows the agent to work with complex data structures like directions, geocoding results, and map features more effectively.

Getting Started

  1. Set your Mapbox access token:

    export MAPBOX_ACCESS_TOKEN=your_token_here

  2. Update the path to your node and the MCP server in the script:

    • If you want to use NPM version:

      server_parameters = StdioServerParameters(
   command=<YOUR_NPX_PATH>,
   args=["-y", "@mapbox/mcp-server"],
   env={
     "MAPBOX_ACCESS_TOKEN": os.environ["MAPBOX_ACCESS_TOKEN"]
   }
)

    • If you want to use the local code version from this repository:

      server_parameters = StdioServerParameters(
   command=<YOUR_NODE_PATH>,
   args=["/YOUR_PATH_TO_REPOSITORY/dist/esm/index.js"],
   env={
     "MAPBOX_ACCESS_TOKEN": os.environ["MAPBOX_ACCESS_TOKEN"]
   }
)

  3. Choose your preferred model by setting the chosen_inference variable

  4. Run the example:

    python smolagents_example.py

The example asks the agent how long it takes to drive from Big Ben to the Eiffel Tower, demonstrating how the agent can use Mapbox's tools to provide a meaningful answer.

If everything works well you can expect output that looks like this: Example of agent response

Customization

You can modify the question at the end of the script to test different location-based queries:

agent.run("Your location-based question here")

Learn More

For more information about: