Build a Working MCP Server in 15 Lines — Hosted Free

What an MCP server actually is

An MCP server is a small backend that exposes your own tools to any AI assistant — Claude, Cursor, Codex, ChatGPT. You write the tool (a few lines of code); the AI sees it in its tool list and calls it during a chat when it fits, with no prompt engineering on your side.

The catch with most "build an MCP server" tutorials: you end up writing OAuth, a database, and hosting before you can write the actual tool. Hatchable hands those to you so you only write the part that matters.

The whole tool, in about 15 lines

This is a complete, working tool that lets your AI save thoughts to your own private notebook from any chat. Nothing else — no auth code, no framework config, no hosting glue.

// Save a thought to your private notebook from any AI chat.
export default {
  access: 'admin',                  // only you can call this
  inputSchema: {
    type: 'object',
    required: ['text'],
    properties: {
      text: { type: 'string', description: 'what to remember' },
      tag:  { type: 'string', description: 'optional tag' },
    },
  },
  handler: async ({ text, tag }, ctx) => {
    const { rows } = await ctx.db.query(
      'insert into thoughts (text, tag) values ($1, $2) returning id',
      [text, tag || null]
    )
    return { ok: true, id: rows[0].id }
  },
}

Plus a one-line migration so the thoughts table exists:

create table thoughts (
  id serial primary key,
  text text not null,
  tag  text,
  created_at timestamptz default now()
);

That is the entire thing. OAuth, hosting, Postgres, and the MCP wire protocol are already handled by the platform.

What this looks like in a chat

Once it is deployed, your AI sees remember in its tool list and uses it when it fits — no prompt engineering required:

What you could build

Once you have one tool live, more take minutes. A few patterns that show up again and again:

Each is one file in mcp/. Add a tool, your AI sees it on the next session.

What is wired up for free

You would otherwise be writing all of this before you wrote the first tool:

• Live, authenticated remote MCP server at {your-slug}.hatchable.site/mcp
• OAuth built in — per-account auth, no token plumbing in your tool code
• Postgres database, file storage, and scheduled jobs — no config
• Sub-second tool updates — drop a file, it is callable on the next request
• Free forever for personal use. Custom domain on Builder ($12/mo flat, covers every project)

And unlike a local MCP server, the URL works when your laptop is closed. Your other AI in another chat can call it. Your friend can call it. It survives.

Connect any MCP client

Once your server is live, paste its URL into anything that speaks MCP Streamable HTTP:

Your tools appear automatically. Your code stays yours — real, portable Node + Postgres source with no proprietary lock-in.