LLNL/napari-mcp

Napari-Mcp

A lightweight napari plugin that exposes the viewer over MCP (Message-Control Protocol) via a Python socket server. Built on top of FastMCP, it lets external MCP-speaking clients—such as autonomous AI agents running on Claude or OpenAI—call napari’s public API remotely.

🔧 Requirements

📦 Napari Installation

python -m pip install "napari[all]"

Install Socket Server Plugin

cd napari-mcp/src/napari_socket
pip install -e .

Install MCP tools in your MCP Client

e.g. For Claude Desktop, go to Developer->Open App Config File and add the below snippet to "mcpServers"

"Napari": {
      "command": ".../python.exe",
      "args": [                        
        ".../napari-mcp/src/napari_mcp/napari_mcp_server.py"
      ],
      "env": {}
    }

🚀 Getting Started

1. Launch napari:

   napari
   

2. Choose Plugins → Socket Server → Start Server. You’ll see something like:

   Listening on 127.0.0.1:64908
   

Interactive Testing

For interactive testing and exploration, use the Jupyter notebook:

cd tests
jupyter notebook test_napari_manager_socket.ipynb

📊 Evaluation

The eval/ directory contains evaluation tools and configurations for testing the MCP server with AI agents.

MCP Client Evaluation

The general_mcp_client.py provides a comprehensive MCP client that supports:

• Multiple LLM providers - Claude, OpenAI, and LiteLLM-compatible endpoints
• Image support - Handles image inputs and outputs for both providers
• Tool execution - Processes MCP tool calls and formats responses
• Error handling - Robust error handling and retry logic

Automated Testing with Promptfoo

Use the test_general.yaml configuration to run automated evaluations:

cd eval
promptfoo eval -c test_general.yaml

This evaluates:

• File loading - Loading TIF files into napari
• Layer management - Checking layer existence and properties
• Screenshot capture - Taking and verifying screenshots
• LLM rubric scoring - AI-powered evaluation of task completion

Evaluation Examples

The eval/eval_examples/ directory contains sample data for evaluation:

• Multi-channel TIF files for testing complex data loading
• Time series data for testing temporal operations

Evaluation Configuration

The evaluation setup supports:

• Custom LLM endpoints - Configure your own API endpoints
• Model selection - Choose different LLM models for evaluation
• Caching control - Enable/disable result caching
• Concurrent execution - Control parallel test execution

Authors

Napari_MCP was created by Haichao Miao (miao1@llnl.gov) and Shusen Liu (liu42@llnl.gov)

License

Napari_MCP is distributed under the terms of the BSD 3‑Clause with Commercial License.

LLNL-CODE-2011142