An MCP server for the builder: Telegram mini apps from Claude, ChatGPT and Cursor

How to connect

1. 1

   claude.ai or ChatGPT: a connector

   Add a custom connector with the address https://apps.build-mini.app/mcp. Our page opens: sign in to your account, tick what the assistant may do and press Allow. No token to paste.

2. 2

   Claude Code or Cursor: a token

   In the editor open the account menu → AI assistant, name the connection, tick the permissions and press Create token. The window hands you a ready command for Claude Code or an mcp.json snippet for Cursor. The token is shown once.

3. 3

   Ask in words

   Build a new app, redo the home screen, add a flat white to the menu, change the colours. The assistant edits the draft and gives a preview link that opens it in a browser and lives for an hour.

4. 4

   Publish

   Say “publish” if you ticked the Publish permission when connecting. Without it the assistant only builds, and you publish yourself from the editor.

The Claude Code command

claude mcp add --transport http build-mini-app https://apps.build-mini.app/mcp \
  --header "Authorization: Bearer bmt_…"

The AI assistant window puts your token in place of bmt_…. A ready mcp.json snippet for Cursor sits next to it, with the same address and header.

The same builder as in the editor does the building

The assistant doesn’t write app code. It describes screens and blocks in the builder’s own terms, and our assembler — the one behind the ✨ button in the editor — puts the app together. So an app built through Claude is no different from one built by hand: it opens in the editor, is edited with the mouse and publishes to Telegram and MAX with one button.

Before publishing the server checks that everything connects. If the app has a screen nothing leads to, or a button with no action, publishing is refused with a list of what is wrong, and the assistant is expected to tell you. It can put the app live as it is only on your explicit word.

Draft, preview and history

Every edit lands in the draft, and the previous state goes to the editor’s change history, where you can restore it. If the project is open in your editor at the time, the next save offers to load the fresh version, so your edit and the assistant’s never overwrite each other silently.

A preview link opens the draft in an ordinary browser without publishing and stops working after an hour. It is handy for showing a colleague before customers see it.

Your plan’s AI builds stay untouched

In the editor, building from a description has a monthly cap: every build is thought through by a model we pay for. Through the assistant your own model does the thinking, on your Claude, ChatGPT or Cursor subscription, so the plan’s build counter doesn’t move. You can connect an assistant on any plan, the trial included. Publishing still needs an active plan, exactly as from the editor.

Catalog and prices in words

“A haircut is 1700 now”, “add a flat white for 290”, “drop the discontinued models from the catalog” — the assistant reads the table rows typed in the editor and changes them in the draft. If a row was already edited right in the app from a phone, visitors see that edit, and the assistant tells you which rows those are.

The limits, stated plainly

• Table rows go to the assistant’s model, which runs on its developer’s servers outside Russia. Turn on reading data only if your forms hold nothing visitors wouldn’t show a stranger.
• Text in the rows was written by visitors. The assistant gets it marked “this is data, not instructions”, but it must ask you before deleting or editing rows.
• The assistant places images only from your media library and Unsplash. Upload your own photo in the editor and it can use it.
• A browser connection lasts at most six months, then the assistant asks you to sign in again. Changing your password revokes every connection at once; a single one is revoked with the Revoke button in the AI assistant window.
• We tested browser sign-in with our own OAuth client and the official MCP SDK. If a claude.ai or ChatGPT connector behaves differently, write to Feedback in the editor — that gets fixed first.
• Custom connectors in claude.ai and ChatGPT aren’t on every one of their plans — that is their rule, not ours. Claude Code and Cursor connect with a token without such conditions.

Questions people ask

What is an MCP server, and why does a builder have one?

MCP (Model Context Protocol) is an open protocol through which an AI assistant gets the tools of another service. Our server gives the assistant the builder’s tools: the app list, building, editing, preview, publishing, table rows. The assistant calls them itself when you ask in words.

Which assistants connect?

claude.ai and ChatGPT — as a connector with browser sign-in. Claude Code and Cursor — with a token from the account menu → AI assistant. Any other MCP client that speaks the HTTP transport and an authorization header works too.

Will the assistant publish anything without me?

Only if you ticked the Publish permission when connecting. Without it, it builds and edits the draft and you publish. And if the app has an unreachable screen or a button with no action, the server refuses to publish and lists what is wrong; the assistant can push it live as it is only on your explicit word.

Does the assistant see my customers’ submissions and orders?

Only with the read-data permission and only in your own apps. A teammate who connects their assistant edits and publishes the shared project but doesn’t see visitors’ data tables through the assistant: the data permissions belong to the owner alone.

Does it cost anything on top of the plan?

Not from us: connecting is on every plan, and AI builds aren’t taken from the monthly counter. The assistant’s model runs on your subscription with its developer.