Introduction to the ESP32-S31 with Vision: Tooling: Set up the ESP-Vision MCP server (optional)

Connect your coding agent to verified vision APIs

#

The Model Context Protocol (MCP) gives an AI coding assistant access to tools and documentation outside its built-in knowledge. See the official ESP-Vision MCP page for service details. In this section, you will connect a compatible client to the endpoint https://mcp.vision.espressif.com.

The connection helps the assistant verify APIs, chip capabilities, examples, and concepts before generating code.

This workshop uses Cursor. You can follow the same workflow with another agent that supports remote MCP servers, such as:

• Visual Studio Code with GitHub Copilot agent mode
• Claude Code
• Claude Desktop
• Another MCP-compatible IDE or coding agent

ESP-Vision MCP server

#

The server can:

• Report modules available on esp32s31, esp32p4, and esp32s3
• Check whether a symbol is available for a selected target
• Return authoritative signatures for ESP-Vision and MicroPython APIs
• Find and return runnable examples
• Explain camera, image, inference, and model workflows

This prevents a common failure mode: generating desktop Python code or a board API that is not present in the ESP-Vision firmware.

Configure your agent

#

Choose the tab for the agent or application you are using.

{
"mcpServers": {
  "esp-vision-mcp": {
    "url": "https://mcp.vision.espressif.com"
  }
}
}

{
"servers": {
  "esp-vision-mcp": {
    "type": "http",
    "url": "https://mcp.vision.espressif.com"
  }
}
}

claude mcp add --transport http esp-vision-mcp \
https://mcp.vision.espressif.com

claude mcp list

Test the integration

#

Ask your assistant:

Use the ESP-Vision MCP server to list the modules supported by esp32s31.
Do not answer from memory.

The answer should include the importable modules sensor, image, display, espdl, and tflite. ImageIO belongs to the image API and is accessed as image.ImageIO; it is not a separate imageio module.

Then ask:

Check the ESP-Vision APIs available on esp32s31 and write a bounded camera
test. Print VISION_TEST:OK only after a frame has been captured.

Review the result:

• It should import sensor, not a desktop camera package.
• It should call sensor.reset() before capture.
• It should use a supported frame size and pixel format.
• It should call sensor.shutdown() during cleanup.
• It should not use pip, requirements.txt, subprocess, or a PC filesystem path.

If the response contains unsupported code, explicitly ask the assistant to check each symbol with the MCP server and regenerate the script.

Complete working configuration

#

{
  "mcpServers": {
    "esp-vision-mcp": {
      "url": "https://mcp.vision.espressif.com"
    }
  }
}

{
  "servers": {
    "esp-vision-mcp": {
      "type": "http",
      "url": "https://mcp.vision.espressif.com"
    }
  }
}

claude mcp add --transport http esp-vision-mcp \
  https://mcp.vision.espressif.com

What you learned

#

In this section, you:

• Connected Cursor, Visual Studio Code, Claude Code, or another client to the ESP-Vision MCP server
• Learned what information the MCP server provides
• Verified ESP32-S31 modules and APIs through tools rather than model memory
• Recognized desktop Python patterns that do not work in device-side MicroPython

Next step

#

Assignment 1: LED blink with ESP-Vision

Return to the workshop main page