MindBridge MCP Server

Instructions

Implementation Reference

• src/server.ts:32-79 (handler)

  The main execution handler for the getSecondOpinion tool. Validates provider and model, fetches the provider instance, calls its getResponse method with input params, and returns formatted LLM response or error.

  async (params) => {
    try {
      // Validate provider exists
      const providerName = params.provider.toLowerCase();
      if (!this.providerFactory.hasProvider(providerName)) {
        const availableProviders = this.providerFactory.getAvailableProviders();
        throw new Error(
          `Provider "${params.provider}" not configured. Available providers: ${availableProviders.join(', ')}`
        );
      }
  
      const provider = this.providerFactory.getProvider(providerName)!;
  
      // Validate model exists for provider
      if (!provider.isValidModel(params.model)) {
        const availableModels = provider.getAvailableModels();
        throw new Error(
          `Model "${params.model}" not found for provider "${params.provider}". Available models: ${availableModels.join(', ')}`
        );
      }
  
      // Check reasoning effort compatibility
      if (params.reasoning_effort && !provider.supportsReasoningEffort()) {
        console.warn(
          `Warning: Provider "${params.provider}" does not support reasoning_effort parameter. It will be ignored.`
        );
      }
  
      // Get response from provider
      const result = await provider.getResponse(params);
  
      if (result.isError) {
        return {
          content: [{ type: 'text', text: `Error: ${result.content[0].text}` }],
          isError: true
        };
      }
  
      return {
        content: result.content
      };
    } catch (error) {
      return {
        content: [{ type: 'text', text: `Error: ${error instanceof Error ? error.message : 'An unknown error occurred'}` }],
        isError: true
      };
    }
  }

• src/server.ts:29-80 (registration)

  Registration of the getSecondOpinion tool in the MCP server using this.tool(), providing name, description, Zod schema, and inline handler function.

  this.tool('getSecondOpinion',
    'Get responses from various LLM providers',
    GetSecondOpinionSchema.shape,
    async (params) => {
      try {
        // Validate provider exists
        const providerName = params.provider.toLowerCase();
        if (!this.providerFactory.hasProvider(providerName)) {
          const availableProviders = this.providerFactory.getAvailableProviders();
          throw new Error(
            `Provider "${params.provider}" not configured. Available providers: ${availableProviders.join(', ')}`
          );
        }
  
        const provider = this.providerFactory.getProvider(providerName)!;
  
        // Validate model exists for provider
        if (!provider.isValidModel(params.model)) {
          const availableModels = provider.getAvailableModels();
          throw new Error(
            `Model "${params.model}" not found for provider "${params.provider}". Available models: ${availableModels.join(', ')}`
          );
        }
  
        // Check reasoning effort compatibility
        if (params.reasoning_effort && !provider.supportsReasoningEffort()) {
          console.warn(
            `Warning: Provider "${params.provider}" does not support reasoning_effort parameter. It will be ignored.`
          );
        }
  
        // Get response from provider
        const result = await provider.getResponse(params);
  
        if (result.isError) {
          return {
            content: [{ type: 'text', text: `Error: ${result.content[0].text}` }],
            isError: true
          };
        }
  
        return {
          content: result.content
        };
      } catch (error) {
        return {
          content: [{ type: 'text', text: `Error: ${error instanceof Error ? error.message : 'An unknown error occurred'}` }],
          isError: true
        };
      }
    }
  );

• src/types.ts:16-35 (schema)

  Zod schema defining the input structure for the getSecondOpinion tool, including required fields like prompt, provider, model, and optional LLM-specific parameters.

  export const GetSecondOpinionSchema = z.object({
    prompt: z.string().min(1),
    provider: LLMProvider,
    model: z.string().min(1),
    systemPrompt: z.string().optional().nullable(),
    temperature: z.number().min(0).max(1).optional(),
    maxTokens: z.number().positive().optional().default(1024),
    reasoning_effort: z.union([ // Primarily for OpenAI o-series
      z.literal('low'),
      z.literal('medium'),
      z.literal('high')
    ]).optional().nullable(),
    // Add other potential parameters if needed based on updated APIs
    top_p: z.number().min(0).max(1).optional(),
    top_k: z.number().positive().optional(),
    stop_sequences: z.array(z.string()).optional(),
    stream: z.boolean().optional(), // For Google Gemini
    frequency_penalty: z.number().min(-2.0).max(2.0).optional(), // For OpenAI
    presence_penalty: z.number().min(-2.0).max(2.0).optional() // For OpenAI
  });