Create client-llm.ts

Author: rohitpackt

07 - building clients/solutions/typescript/src/client-llm.ts

@@ -0,0 +1,162 @@
+import { Client } from "@modelcontextprotocol/sdk/client/index.js";
+import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
+import OpenAI from "openai";
+import { z } from "zod"; 
+
+import readline from 'node:readline';
+import process from "node:process";
+
+const transport = new StdioClientTransport({
+  command: "node",
+  args: ["./build/index.js"]
+});
+
+const client = new Client(
+  {
+    name: "example-client",
+    version: "1.0.0"
+  }
+);
+
+const openai = new OpenAI({
+    baseURL: "https://models.github.ai/inference", 
+    apiKey: process.env.GITHUB_TOKEN,
+});
+
+const read = readline.createInterface({
+    input: process.stdin,
+    output: process.stdout
+});
+
+// 0. Add convert function
+function toLLMTool(tool: {
+    name: string;
+    description?: string;
+    inputSchema: any;
+}) {
+        // Create a zod schema based on the input_schema
+        const schema = z.object(tool.inputSchema);
+    
+        return {
+        type: "function" as const, // Explicitly set type to "function"
+        function: {
+            name: tool.name,
+            description: tool.description,
+            parameters: {
+                type: "object",
+                properties: tool.inputSchema.properties,
+                required: tool.inputSchema.required,
+            },
+        },
+        };
+}
+
+// function that calls MCP tool based on LLm response
+async function callTools(
+    tool_calls: OpenAI.Chat.Completions.ChatCompletionMessageToolCall[],
+    toolResults: any[]
+    ): Promise<void> {
+    for (const tool_call of tool_calls) {
+        const toolName = tool_call.function.name;
+        const args = tool_call.function.arguments;
+
+        console.log(`Calling tool "${toolName}" with args ${JSON.stringify(args)}`);
+
+
+        // 2. Call the server's tool 
+        const toolResult = await client.callTool({
+            name: toolName,
+            arguments: JSON.parse(args),
+        });
+
+        console.log("\nTool result: ", toolResult);
+
+        // 3. Do something with the result
+        // TODO  
+
+    }
+}
+
+async function main() {
+    await client.connect(transport);
+
+    // List tools
+    const tools = await client.listTools();
+
+    
+
+    let toolDescriptions = "";
+    const llmTools = [];
+
+    for(let tool of tools.tools) {
+        toolDescriptions += `${tool.name}, ${tool.description}\n`;
+
+        // 1. convert this response to LLM tool
+        llmTools.push(toLLMTool(tool));
+    }
+
+    console.log("Welcome to E-shopping server");
+    let keepRunning = true;
+
+    while(keepRunning) {
+
+        let command = await getInput("Provide user prompt (type h for help or quit to stop): ");
+        console.log(command);
+
+        if(command == "quit") {
+            break;
+        } else if(command == "h") {
+            console.log("FEATURES");
+            console.log(toolDescriptions);
+        } else {
+
+            // 2. take the command response as a user prompt to the LLM
+            const messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[] = [
+                {
+                    role: "user",
+                    content: command,
+                },
+            ];
+
+            // 3. make LLM call
+            console.log("Querying LLM: ", messages[0].content);
+
+            let response = openai.chat.completions.create({
+                model: "gpt-4o-mini",
+                max_tokens: 1000,
+                messages,
+                tools: llmTools,
+            });    
+            // 4. read LLM response and call suitable MCP Server tool if any..
+            let results: any[] = [];
+        
+            // 1. Go through the LLM response,for each choice, check if it has tool calls 
+            for (const choice of (await response).choices) {
+                const message = choice.message;
+                if (message.tool_calls) {
+                    console.log("Making tool call")
+                    await callTools(message.tool_calls, results);
+                }
+            }
+        }
+
+    }
+    read.close();
+    
+    console.log("Disconnected, Bye!");
+    return;
+}
+
+function getInput(query: string): Promise<string> {
+    return new Promise(resolve => {
+        read.question(query, (answer: string) => {
+            resolve(answer);
+        });
+    });
+}
+
+main().catch((error) => {
+    console.error("Error: ", error);
+    
+});
+