commitd
Get started

MCP, explained

A menu of actions your customer's assistant may use.

An MCP server is a small list of things an AI assistant is allowed to do on your behalf: look up an order, book a slot, read your programme. You write the list. The assistant your customer already uses does the talking.

Build one How to build it
a conversationcoach.commitd.ai/mcp
client    What is on the plan for week 4?
assistant → tools/call programme_week { week: 4 }
assistant Week 4 is a lighter week: two short
          runs and one long walk. Your coach left
          a note: "Sleep comes first this week."
client    Book a check-in for Thursday.
assistant → tools/call book_session { date: "2026-11-05" }
assistant Done. Thursday 5 November, 10:00.
tools · resources · prompts

Two stories

01What it looks like in practice

a coach

Sanne runs a 12-week running programme

Her clients used to get a PDF and a WhatsApp group. Now each client connects her MCP server to Claude and talks to it like a personal assistant that knows the programme.

  1. programme_week

    Returns the plan for a given week, with Sanne's notes. Read-only, safe for everyone.

  2. weekly_checkin

    Stores how the week went: energy, sleep, pain. Sanne sees it in her dashboard the same evening.

  3. book_session

    Books a call in a free slot. Each client only sees and books their own. Login makes that possible.

a webshop

A bike shop with 4,000 orders a year

Half their support email was "where is my order" and "can I return this". Their MCP server answers both, and the product FAQ, inside whatever assistant the customer prefers.

  1. order_status

    Looks the order up in the shop through a connection. Login-only, so a customer sees only their own orders.

  2. start_return

    Opens a return, emails the label. Writes to the database and a job notifies the warehouse.

  3. product_faq

    Answers from a document the shop keeps up to date. Public: anyone may ask.

What you get

02Everything around the tools

You write the tools. We run everything an MCP server needs to be safe to hand to real people.

address

A hosted URL

your-name.commitd.ai/mcp, with a certificate, on fast edge servers near your users. Your own domain on paid plans.

login

Login that assistants understand

Claude, Cursor and friends expect OAuth. We handle it on your app's address. Your users click a link in their email, once.

people

Invites and roles

Add users by email, give them roles. A tool can say ctx.auth.require('client') and that is the whole access check.

history

Versions and rollback

Every deploy is numbered and kept. Changed a tool and regret it? Roll back in one click. No redeploy.

How to build it

03Three ways, all short

You do not need to know what JSON-RPC is. You need to know what you want the assistant to be able to do.

  1. Ask your AI tool to write it

    Copy our prompt into Claude or ChatGPT and describe your tools in plain words. It answers with one file that follows our rules.

  2. Paste and deploy

    Create a project in the dashboard, turn on MCP and the blocks you need, paste the file, press Deploy. The dashboard shows your address and a test console.

  3. Or start from a template

    Pick the coach template or the shop template, change the texts and tables, deploy. Ten minutes, honestly.

  4. Share the address

    Send your-name.commitd.ai/mcp to your clients. Their assistant handles the rest, including login.

Connecting

04What your clients do

Claude.ai and Claude Desktop
Settings → Connectors → Add custom connector
URL: https://your-name.commitd.ai/mcp
Claude Code
claude mcp add --transport http your-name \
  https://your-name.commitd.ai/mcp
Cursor and others · mcp.json
{ "mcpServers": {
    "your-name": { "url": "https://your-name.commitd.ai/mcp" }
} }
Any client with a token
Authorization: Bearer <token>
Tokens: https://your-name.commitd.ai/_auth/tokens

Questions

05Before you start

Is it safe?

The assistant can only do what your tools allow, nothing more. Each app runs on its own, with its own database. Login-only servers require a signed-in user for every call, and your code sees who that is. Store credentials for outside systems as encrypted app secrets and read them through ctx.env.

Who can call it?

You choose per server. Public: anyone with the address. Login-only: people you invited, or anyone who signs in if you allow self sign-up. Roles let you separate clients from staff.

What does it cost?

Pro is €19 a month for up to 10 apps or MCP servers, Business €79 for 50. Both include login. See pricing.

Can I use my own domain?

Yes, on Pro and Business. Point a subdomain like assistant.yourbrand.com at us and your MCP server answers there.

Which assistants work?

Anything that speaks MCP over HTTP: Claude.ai, Claude Desktop, Claude Code, Cursor, Windsurf, ChatGPT connectors and many more. Login uses standard OAuth, so new clients work without changes on your side.

Do I need to write code?

You need one file. Most people let their AI tool write it from our prompt, or start from a template and change the texts. The dashboard has a test console so you can try every tool before sharing it.

Get started

Your clients' assistant, briefed.

Build one Copy the prompt