h3-mcp-tools

Minimal MCP server built with H3 v2 (beta) as a dedicated app or subapp.

Features

• Built with H3 v2 (beta) for high performance and runtime agnosticity (it can be mounted as a nested-app).
• It is based on the JSON-RPC protocol to accept commands and return results.
• Tools:
  • Simple Tool registration and invocation with built-in input validation via Standard-Schema.
  • Automatic JSON Schema generation for supported validation libraries.
  • Tool listing method to retrieve all statically registered tools.
  • Custom handlers for listing and calling dynamic tools.
• Resources:
  • Support for static resources definition (including batch definition).
  • Support for resource templates.
  • Custom handlers for listing and reading dynamic resources (like fetched from a remote storage).
• Streaming:
  • Streaming responses via ReadableStream with a simple utility function.

[!WARNING]
Project currently under heavy development. The main scope of this project is to provide a simple MCP server with minimal dependencies, alongside showcasing H3 v2 (beta). If you are looking for more capabilities out of the box I suggest you to look for the official MCP TS SDK or the great TMCP.

TODO

• Support for prompt templates definition
• Completion API for resources
• Built-in Streamable HTTP notifications hooks

Usage

Install the package:

# ✨ Auto-detect (supports npm, yarn, pnpm, deno and bun)
npx nypm install h3-mcp-tools

H3MCP App

The easiest way to get started is to use the H3MCP class, which extends H3.

import { H3MCP } from "h3-mcp-tools"; // or from CDN via "https://esm.sh/h3-mcp-tools"
import { serve } from "h3";
import * as v from "valibot";

const app = new H3MCP({
  name: "My MCP Server",
  version: "1.0.0",
  description: "A sample MCP server built with H3",
});

app.tool(
  {
    name: "echo",
    description: "Echoes back the input",
    schema: v.object({
      input: v.string(),
    }),
  },
  async ({ input }) => {
    return { output: `You said: ${input}` };
  },
);

serve(app);

Tools

You can define tools using the tool method on the H3MCP instance.

Requesting a Tool Call

Do a POST request to the /mcp endpoint with the following body:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "echo",
    "arguments": {
      "input": "hello from h3"
    }
  }
}

You should receive a response like this:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "output": "You said: hello from h3"
  }
}

Requesting Tool Listing

To list all registered tools, you can make a POST request to the /mcp endpoint with the following body:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/list"
}

You should receive a response like this:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "tools": [
      {
        "name": "echo",
        "description": "Echoes back the input",
        "inputSchema": {
          "$schema": "http://json-schema.org/draft-07/schema#",
          "type": "object",
          "properties": {
            "input": { "type": "string" }
          },
          "required": ["input"]
        }
      }
    ]
  }
}

JSON Schema

Only validation libraries supported by @standard-community/standard-json are automatically evaluated at runtime. For Valibot and Zod users you should also install the related json schema package (@valibot/to-json-schema, zod-to-json-schema).

If your validation library is supported by the @standard-schema/spec but not from the @standard-community/standard-json, you can still use it by providing the jsonSchema option in the tool definition, like this:

app.tool(
  {
    name: "test",
    description: "An example tool that echoes back the input",
    jsonSchema: {
      $schema: "http://json-schema.org/draft-07/schema#",
      type: "object",
      properties: {
        input: { type: "string" },
      },
      required: ["input"],
    },
  },
  async ({ input }) => {
    return { output: `You said: ${input}` };
  },
);

Streaming Responses

While standard SSE (Server-Sent Events) are deprecated in the MCP specification, you can still implement streaming responses using ReadableStream. To make things simpler, you can use the createMcpStream utility function provided by this package.

You can create a streamable response like this:

import { H3MCP, createMcpStream } from "h3-mcp-tools";
import { serve } from "h3";
import * as v from "valibot";

const app = new H3MCP({
  name: "My streamable MCP Server",
  version: "1.0.0",
  description: "A sample MCP server built with H3",
});

app.tool(
  {
    name: "stream",
    description: "Streams the current time every second",
    schema: v.optional(
      v.object({
        maxSeconds: v.optional(v.number()),
      }),
    ),
  },
  ({ maxSeconds } = {}, event, { id }) => {
    let count = 0;
    const max = maxSeconds ?? 10;

    const stream = createMcpStream(event, {
      async start(controller) {
        for (let i = 0; i < max; i++) {
          if (count < max) {
            controller.enqueue(
              new TextEncoder().encode(`Time: ${new Date().toISOString()}\n`),
            );
            count++;
          }
          await new Promise((resolve) => setTimeout(resolve, 1000));
        }
      },
      finalResponse: {
        jsonrpc: "2.0",
        id,
        result: { message: "Stream completed" },
      },
    });

    event.res.headers.set("Content-Type", "text/event-stream");

    return stream;
  },
);

serve(app);

If you include finalResponse in the createMcpStream options, it will be sent at the end of the stream, allowing you to send a complete JSON-RPC response with the jsonrpc, id, and result fields and the stream will be closed automatically.

Resources

You can define static resources using the resource method, which accepts a Resource object. You can also define dynamic resources with custom handlers.

import { H3MCP } from "h3-mcp-tools";
import { serve } from "h3";

const app = new H3MCP({
  description: "A sample MCP server built with H3",
});

app.resource(
  {
    uri: "hello/world",
    name: "Welcome Resource",
    title: "A simple resource that returns a greeting",
    description: "Returns a greeting message",
    mimeType: "test/plain",
  },
  async ({ uri }) => {
    // uri = "hello/world"
    return { text: `Hello, ${uri}!` };
  },
);

serve(app);

Batch Resource Definition

You can also define multiple resources at once using the resource method, which accepts an array of Resource objects and an optional handler, which will be shared among all resources:

const remotelyFetchedResources = [
  {
    uri: "foo",
    name: "Foo Resource",
    text: "This is a Foo resource",
  },
  {
    uri: "bar",
    name: "Bar Resource",
    text: "This is a Bar resource",
  },
];

app.resource(remotelyFetchedResources, ({ uri }) => {
  return { text: `This is a ${uri} resource` };
});

Custom Handlers

You can override the default behavior for listing and calling tools, or listing and reading resources, by providing your own handlers. This becomes useful when you want to dynamically manage tools or resources to other sources, like a database or a remote API.

// Override tools/list
app.toolsList(({ tools }) => {
  return {
    tools: [
      ...tools, // Include statically defined tools
      {
        name: "custom-listed-tool",
        description: "A tool added via a custom handler",
      },
    ],
  };
});

// Override tools/call for non-statically defined tools
app.toolsCall(async (params) => {
  if (params.name === "custom-handled-tool") {
    return { result: "Handled by custom toolsCall" };
  }
  // Fallback to default behavior by returning nothing (undefined)
});

Standalone Handler

For more advanced use cases, you can use defineMcpHandler to create a standalone handler that can be used as a sub-app.

import { defineMcpHandler } from "h3-mcp-tools";
import { serve, H3 } from "h3";
import * as v from "valibot";

const mcpHandler = defineMcpHandler({
  serverInfo: {
    name: "My MCP Server",
    version: "1.0.0",
  },
  tools: [
    {
      definition: {
        name: "echo",
        description: "Echoes back the input",
        schema: v.object({
          input: v.string(),
        }),
      },
      handler: async ({ input }) => {
        return { output: `You said: ${input}` };
      },
    },
  ],
});

const app = new H3().all("/mcp", mcpHandler);

serve(app);

Notes

• SSE have been marked as deprecated since MCP spec "2025-03-26", instead you should use ReadableStream for streaming responses with Content-Type: text/event-stream header. Please also note that you should send a full JSON-RPC response at the end of the stream, with the jsonrpc, id and result fields.

Development

• Clone this repository
• Install latest LTS version of Node.js
• Enable Corepack using corepack enable
• Install dependencies using pnpm install
• Run interactive tests using pnpm dev

Credits

Inspired by:

• the official ModelContextProtocol SDK
• Paolo Ricciuti's work on tmcp

License

Published under the MIT license. Made by community 💛

🤖 auto updated with automd