MCP Server

Once we’ve covered Function Calling, Real-time UI but also how to do real-time polling, it’s time to add the last bit of the puzzle and proof that your API is an MCP if it’s built with Vovk.ts. For that we’re going to use @vercel/mcp-adapter package that allows us to easily create an MCP server on Nest.js.

The createLLMTools function accepts a resultFormatter option that accepts a function to format the result of the tool execution that’s coming back to the LLM, but also a string with a special value mcp that formats the result in a way that MCP server expects. This is all we need to do to make our tools output MCP-compatible and return the results in the following format:


{
    content: [
      {
        type: 'text',
        text: 'Result text',
      },
    ],
}


import { createLLMTools } from "vovk";
import UserController from "@/modules/user/UserController";
import TaskController from "@/modules/task/TaskController";
 
const { tools } = createLLMTools({
  modules: {
    UserController,
    TaskController,
  },
  resultFormatter: "mcp",
});

There is where custom operation properties such as x-tool-successMessage and x-tool-errorMessage will be useful to make the output more LLM-friendly.


import { prefix, get, put, post, del, operation } from "vovk";
import UserService from "./UserService";
import { z } from "zod";
import { BASE_FIELDS } from "@/constants";
import { UserSchema } from "../../../prisma/generated/schemas";
import { withZod } from "@/lib/withZod";
 
@prefix("users")
export default class UserController {
  @operation({
    summary: "Update user",
    description:
      "Updates an existing user with the provided details, such as their email or name.",
    "x-tool-successMessage": "User updated successfully",
    "x-tool-errorMessage": "Failed to update user",
  })
  @put("{id}")
  static updateUser = withZod({
    body: UserSchema.omit(BASE_FIELDS).partial(),
    params: UserSchema.pick({ id: true }),
    handle: async ({ vovk }) =>
      UserService.updateUser(vovk.params().id, await vovk.body()),
  });
}

This will turn the text property of the output to “User updated successfully” or “Failed to update user” depending on the result of the operation, concatenated with the actual serialized result or error message.


{
    content: [
      {
        type: 'text',
        text: 'User updated successfully\nResult: { id: 1, email: "user@example.com" }',
      },
    ],
}

Now create the route file as per the adapter documentation .

src/app/api/mcp/route.mdx


import { createMcpHandler } from "@vercel/mcp-adapter";
import { createLLMTools } from "vovk";
import UserController from "@/modules/user/UserController";
import TaskController from "@/modules/task/TaskController";
import { convertJsonSchemaToZod } from "zod-from-json-schema";
import { mapValues } from "lodash";
 
const { tools } = createLLMTools({
  modules: {
    UserController,
    TaskController,
  },
  resultFormatter: "mcp",
  onError: (e) => console.error("Error", e),
});
 
const handler = createMcpHandler(
  (server) => {
    tools.forEach(({ name, execute, description, parameters }) => {
      server.tool(
        name,
        description,
        mapValues(parameters?.properties ?? {}, convertJsonSchemaToZod),
        execute,
      );
    });
  },
  {},
  { basePath: "/api" },
);
 
export { handler as GET, handler as POST, handler as DELETE };

As you can see, we convert the JSON Schema parameters of each tool to Zod schemas using zod-from-json-schema package, as MCP adapter expects Zod schemas. This will work for all available validation libraries. But if you use Zod, you can also use models property of a tool to get the Zod schema directly, as shown below:


// ...
const handler = createMcpHandler(
  (server) => {
    tools.forEach(({ name, execute, description, models }) => {
      server.tool(
        name,
        description,
        models,
        execute,
      );
    });
  },
  {},
  { basePath: "/api" },
);
// ...

Note that the models property is only available for callable handlers but not for RPC modules.

As a small reminder, you can also use 3rd party OpenAPI mixins to add more tools to your MCP server.


import { GithubIssuesRPC } from "vovk-client";
// ...
const githubOptions = {
    init: {
        headers: {
            Authorization: `Bearer ${process.env.GITHUB_TOKEN}`,
            "X-GitHub-Api-Version": "2022-11-28",
        },
    },
};
const { tools } = createLLMTools({
    modules: {
        UserController,
        TaskController,
        GithubIssuesRPC: [GithubIssuesRPC, githubOptions],
    },
});
// ...