MCP Elicitations Demo Server

import { z } from "zod"; import { zodToJsonSchema } from "zod-to-json-schema"; import { Server } from "@modelcontextprotocol/sdk/server/index.js"; import { requestSampling, ModelPreferences } from "../lib/sampling.js"; const ModelHintSchema = z.object({ name: z .string() .describe("Model name hint (e.g., 'claude-3-sonnet', 'gpt-4')"), }); const ModelPreferencesSchema = z.object({ hints: z .array(ModelHintSchema) .optional() .describe("Ordered list of preferred model hints"), costPriority: z .number() .min(0) .max(1) .optional() .describe("Priority for cost optimization (0-1, higher = prefer cheaper)"), speedPriority: z .number() .min(0) .max(1) .optional() .describe("Priority for speed optimization (0-1, higher = prefer faster)"), intelligencePriority: z .number() .min(0) .max(1) .optional() .describe("Priority for intelligence (0-1, higher = prefer more capable)"), }); const SampleWithPreferencesSchema = z.object({ prompt: z.string().describe("The prompt to send to the LLM"), systemPrompt: z .string() .optional() .describe("System prompt to guide the LLM's behavior"), maxTokens: z .number() .default(100) .describe("Maximum number of tokens to generate"), modelPreferences: ModelPreferencesSchema.optional().describe( "Model selection preferences" ), }); export const sampleWithPreferencesTool = { name: "sampleWithPreferences", description: "Sample from an LLM with detailed model preferences", inputSchema: zodToJsonSchema(SampleWithPreferencesSchema), handler: async (args: any, request: any, server: Server) => { const validatedArgs = SampleWithPreferencesSchema.parse(args); const { prompt, systemPrompt, maxTokens, modelPreferences } = validatedArgs; const result = await requestSampling( { messages: [ { role: "user", content: { type: "text", text: prompt, }, }, ], systemPrompt, maxTokens, modelPreferences: modelPreferences as ModelPreferences, }, server ); return { content: [ { type: "text" as const, text: `Model used: ${result.model || "unknown"}\nStop reason: ${ result.stopReason || "unknown" }\n\nResponse: ${result.content.text}`, }, ], }; }, };