LibreModel MCP Server

Instructions

Implementation Reference

• src/index.ts:79-109 (handler)

  Handler function for the 'chat' tool that invokes callLlamaServer with user parameters and formats the response as MCP content.

  }, async (args) => {
    try {
      const response = await this.callLlamaServer({
        message: args.message,
        temperature: args.temperature || this.config.defaultTemperature,
        max_tokens: args.max_tokens || this.config.defaultMaxTokens,
        top_p: args.top_p || this.config.defaultTopP,
        top_k: args.top_k || this.config.defaultTopK,
        system_prompt: args.system_prompt || ""
      });
  
      return {
        content: [
          {
            type: "text",
            text: `**LibreModel (Gigi) responds:**\n\n${response.content}\n\n---\n*Tokens: ${response.tokens_predicted} | Model: ${response.model || "LibreModel"}*`
          }
        ]
      };
    } catch (error) {
      return {
        content: [
          {
            type: "text", 
            text: `**Error communicating with LibreModel:**\n${error instanceof Error ? error.message : String(error)}`
          }
        ],
        isError: true
      };
    }
  });

• src/index.ts:71-78 (schema)

  Zod input schema defining parameters for the 'chat' tool.

  inputSchema: {
    message: z.string().describe("Your message to LibreModel"),
    temperature: z.number().min(0.0).max(2.0).default(this.config.defaultTemperature).describe("Sampling temperature (0.0-2.0)"),
    max_tokens: z.number().min(1).max(2048).default(this.config.defaultMaxTokens).describe("Maximum tokens to generate"),
    top_p: z.number().min(0.0).max(1.0).default(this.config.defaultTopP).describe("Nucleus sampling parameter"),
    top_k: z.number().min(1).default(this.config.defaultTopK).describe("Top-k sampling parameter"),
    system_prompt: z.string().default("").describe("Optional system prompt to prefix the conversation")
  }

• src/index.ts:68-109 (registration)

  Registration of the 'chat' tool on the MCP server, including title, description, schema, and handler.

  this.server.registerTool("chat", {
    title: "Chat with LibreModel",
    description: "Have a conversation with LibreModel (Gigi)",
    inputSchema: {
      message: z.string().describe("Your message to LibreModel"),
      temperature: z.number().min(0.0).max(2.0).default(this.config.defaultTemperature).describe("Sampling temperature (0.0-2.0)"),
      max_tokens: z.number().min(1).max(2048).default(this.config.defaultMaxTokens).describe("Maximum tokens to generate"),
      top_p: z.number().min(0.0).max(1.0).default(this.config.defaultTopP).describe("Nucleus sampling parameter"),
      top_k: z.number().min(1).default(this.config.defaultTopK).describe("Top-k sampling parameter"),
      system_prompt: z.string().default("").describe("Optional system prompt to prefix the conversation")
    }
  }, async (args) => {
    try {
      const response = await this.callLlamaServer({
        message: args.message,
        temperature: args.temperature || this.config.defaultTemperature,
        max_tokens: args.max_tokens || this.config.defaultMaxTokens,
        top_p: args.top_p || this.config.defaultTopP,
        top_k: args.top_k || this.config.defaultTopK,
        system_prompt: args.system_prompt || ""
      });
  
      return {
        content: [
          {
            type: "text",
            text: `**LibreModel (Gigi) responds:**\n\n${response.content}\n\n---\n*Tokens: ${response.tokens_predicted} | Model: ${response.model || "LibreModel"}*`
          }
        ]
      };
    } catch (error) {
      return {
        content: [
          {
            type: "text", 
            text: `**Error communicating with LibreModel:**\n${error instanceof Error ? error.message : String(error)}`
          }
        ],
        isError: true
      };
    }
  });

• src/index.ts:297-338 (helper)

  Helper method that constructs the prompt and makes the HTTP POST request to the llama-server /completion endpoint, used by the chat handler.

  private async callLlamaServer(params: {
    message: string;
    temperature: number;
    max_tokens: number;
    top_p: number;
    top_k: number;
    system_prompt: string;
  }): Promise<LlamaCompletionResponse> {
    const prompt = params.system_prompt 
      ? `${params.system_prompt}\n\nHuman: ${params.message}\n\nAssistant:`
      : `Human: ${params.message}\n\nAssistant:`;
  
    const requestBody: LlamaCompletionRequest = {
      prompt,
      temperature: params.temperature,
      n_predict: params.max_tokens,
      top_p: params.top_p,
      top_k: params.top_k,
      stop: this.config.stopSequences,
      stream: false
    };
  
    const response = await fetch(`${this.config.url}/completion`, {
      method: "POST",
      headers: {
        "Content-Type": "application/json"
      },
      body: JSON.stringify(requestBody)
    });
  
    if (!response.ok) {
      throw new Error(`HTTP ${response.status}: ${response.statusText}`);
    }
  
    const data = await response.json() as LlamaCompletionResponse;
    
    if (!data.content) {
      throw new Error("No content in response from llama-server");
    }
  
    return data;
  }