MCP Server Gemini

Instructions

Implementation Reference

• src/enhanced-stdio-server.ts:740-784 (handler)

  The primary handler method for the 'count_tokens' tool. It extracts parameters, calls the underlying Gemini SDK's countTokens method, formats the token count result into an MCP response, and handles errors.

  private async countTokens(id: any, args: any): Promise<MCPResponse> {
    try {
      const model = args.model || 'gemini-2.5-flash';
  
      const result = await this.genAI.models.countTokens({
        model,
        contents: [
          {
            parts: [
              {
                text: args.text
              }
            ],
            role: 'user'
          }
        ]
      });
  
      return {
        jsonrpc: '2.0',
        id,
        result: {
          content: [
            {
              type: 'text',
              text: `Token count: ${result.totalTokens}`
            }
          ],
          metadata: {
            tokenCount: result.totalTokens,
            model
          }
        }
      };
    } catch (error) {
      return {
        jsonrpc: '2.0',
        id,
        error: {
          code: -32603,
          message: error instanceof Error ? error.message : 'Internal error'
        }
      };
    }
  }

• src/enhanced-stdio-server.ts:319-337 (registration)

  Registers the 'count_tokens' tool in the list of available tools, defining its name, description, and input schema for MCP protocol.

    name: 'count_tokens',
    description: 'Count tokens for a given text with a specific model',
    inputSchema: {
      type: 'object',
      properties: {
        text: {
          type: 'string',
          description: 'Text to count tokens for'
        },
        model: {
          type: 'string',
          description: 'Model to use for token counting',
          enum: Object.keys(GEMINI_MODELS),
          default: 'gemini-2.5-flash'
        }
      },
      required: ['text']
    }
  },

• src/utils/validation.ts:93-96 (schema)

  Zod schema definition for validating input parameters of the 'count_tokens' tool (text required, model optional).

  countTokens: z.object({
    text: z.string().min(1, 'Text is required'),
    model: CommonSchemas.geminiModel.optional()
  }),