OpenXAI MCP Server

Instructions

Implementation Reference

• index.js:407-464 (handler)

  The primary handler function that executes the load_dataset tool. It retrieves dataset metadata for supported datasets (german, compas, adult), validates the dataset_name, and returns formatted information including dataset stats and a Python code example for loading the dataset using OpenXAI.

    async loadDataset(datasetName, download = true) {
      const datasetInfo = {
        german: {
          description: 'German Credit dataset loaded successfully',
          features: 20,
          samples: 1000,
          classes: 2,
          task: 'classification'
        },
        compas: {
          description: 'COMPAS Recidivism dataset loaded successfully',
          features: 11,
          samples: 6172,
          classes: 2,
          task: 'classification'
        },
        adult: {
          description: 'Adult Income dataset loaded successfully',
          features: 14,
          samples: 48842,
          classes: 2,
          task: 'classification'
        }
      };
  
      const info = datasetInfo[datasetName];
      if (!info) {
        throw new Error(`Dataset '${datasetName}' not found. Available datasets: ${Object.keys(datasetInfo).join(', ')}`);
      }
  
      const codeExample = `
  # Example usage with OpenXAI:
  from openxai.dataloader import ReturnLoaders
  
  # Load the dataset
  trainloader, testloader = ReturnLoaders(data_name='${datasetName}', download=${download})
  
  # Get a sample from the test dataset
  inputs, labels = next(iter(testloader))
  print(f"Input shape: {inputs.shape}")
  print(f"Labels shape: {labels.shape}")
  `;
  
      return {
        content: [
          {
            type: 'text',
            text: `${info.description}\n\n` +
                  `Dataset: ${datasetName}\n` +
                  `Features: ${info.features}\n` +
                  `Samples: ${info.samples}\n` +
                  `Classes: ${info.classes}\n` +
                  `Task: ${info.task}\n\n` +
                  `Python code example:\n\`\`\`python${codeExample}\`\`\``
          }
        ]
      };
    }

• index.js:56-70 (schema)

  Input schema definition for the load_dataset tool, specifying dataset_name as required string and optional download boolean.

  inputSchema: {
    type: 'object',
    properties: {
      dataset_name: {
        type: 'string',
        description: 'Name of the dataset to load (e.g., german, compas, adult)',
      },
      download: {
        type: 'boolean',
        description: 'Whether to download the dataset if not available locally',
        default: true
      }
    },
    required: ['dataset_name']
  }

• index.js:53-71 (registration)

  Registration of the load_dataset tool in the MCP server's tool list, including name, description, and input schema.

  {
    name: 'load_dataset',
    description: 'Load a specific dataset from OpenXAI',
    inputSchema: {
      type: 'object',
      properties: {
        dataset_name: {
          type: 'string',
          description: 'Name of the dataset to load (e.g., german, compas, adult)',
        },
        download: {
          type: 'boolean',
          description: 'Whether to download the dataset if not available locally',
          default: true
        }
      },
      required: ['dataset_name']
    }
  },

• index.js:258-259 (helper)

  Dispatch case in the CallToolRequestHandler that routes load_dataset calls to the loadDataset method.

  case 'load_dataset':
    return await this.loadDataset(args.dataset_name, args.download);