Let your agents make games: vbgnt is now an MCP server

Until now the only way to make a game on vbgnt was to open the studio and type. As of this week your own AI agent can do it for you. The studio is an MCP server: connect it to Claude, to Claude Code, or to any assistant that speaks the protocol, sign in once, and that agent can create games on your account, follow the builds, tidy your arcade and export the results, from wherever you already talk to it.

This post covers what an MCP server is, what yours can do on vbgnt, how the permission screen works, what it is good for, and the list of things it deliberately cannot do.

What is an MCP server?

The Model Context Protocol is the open standard AI assistants use to reach tools outside themselves. An MCP server is a web address that describes a set of tools ("create a game", "list my exports") in a form an agent can read, and carries them out when asked. Claude, ChatGPT, Cursor and most coding agents can be pointed at one. The agent sees the tools, the user sees a permission screen, and after that the agent calls the tools as part of an ordinary conversation.

vbgnt's server lives at https://api.vbgnt.com/mcp. It exposes 27 tools in four groups, and every one of them acts on the account of the person who approved the connection, never on anyone else's.

What your agent can do

• Games. Create a game from a name, a style and a brief, exactly as you would in the studio. Send follow-up instructions to a game that exists. Watch a build's progress, read the transcript, stop a turn, fork a game, rename it, write its blurb, feature it or hide it. Browse, read and edit the project's files, download the whole project as a zip, or rebuild the web build after an edit.
• Arcade. Check whether an arcade name is free and claim it. Read the arcade's traffic: views, players, plays and playtime, per game and per platform. Reorder the shelf.
• Exports. Start a Windows, macOS, Linux, Android, iOS or web export, follow it, fetch the download links, and push the builds to itch.io when the account has an itch key.
• Account. Ask who is signed in, which arcade the account owns, which models it can build with, and which keys are set.

The tools are the same functions the studio's own buttons call. An agent that creates a game through MCP gets the same cloud machine, the same generated art, the same headless-browser test and the same publish to yourname.vbgnt.com as a person typing the brief.

How to connect

You need a vbgnt account, which for now means an invite. Then:

• Claude.ai. Settings, Connectors, Add custom connector, and paste the server address. Claude opens a vbgnt page in your browser, you sign in and press Allow, and the connector turns green.
• Claude Code. One command in the terminal, then the same sign-in page:

  claude mcp add --transport http vbgnt https://api.vbgnt.com/mcp

• Anything else. Any client that implements the protocol's authorization flow works the same way. There is no API key to paste anywhere: the sign-in page is the only door.

Once connected, ask for a game the way you would in the studio. "Make a two-player online tank duel, top-down, 90s shareware look" is a complete request. The agent calls create_game, the build starts on a cloud machine, and the agent checks back on it until the game is live and hands you the link.

What the permission screen shows

Connecting an agent to an account that can spend build time and publish things to the web deserves a proper permission flow, so we used the same scheme brokerages and banks use for their agent connectors: the server is its own OAuth authorization server, the agent registers itself, and every token is bound to the browser that approved it with a one-time proof.

In practice you see one page. It names the agent that is asking, the account it will act as, and the four permissions it wants: games, arcade, exports, account. An agent can ask for fewer, and a connection only ever gets the tools in the groups it was granted. Deny sends the agent away with nothing. Allow issues a token that expires after an hour and refreshes quietly for as long as the agent stays connected; disconnect the agent on its side and the refresh stops, so the connection dies on its own.

One rule we chose on purpose: an agent can never set your keys. Your Claude token, your OpenAI sign-in, your image and audio keys and your itch.io key are entered in the studio's Settings and nowhere else. An agent can see whether a key exists. It cannot read one, and it cannot add one, because anything pasted into an agent conversation ends up in a transcript somewhere.

What it is good for

The interesting cases are the ones where the game is a step in something larger, not the whole job.

• A story that becomes a game. A family writes a bedtime story with an assistant over a week, then says "make this a small game she can play on the tablet". The agent already knows the characters, the places and the rules of the world, and writes a better brief than a stranger would.
• A teacher's assistant. One conversation, ten short games, each built to a lesson plan the assistant helped write, each hidden from the public arcade until the lesson it belongs to.
• A release manager. A scheduled agent that reads the arcade's traffic every Monday, exports the most played game for every platform, pushes it to itch.io and reports back. No studio tab open.
• A coding agent with taste. Claude Code, working in a Godot project it downloaded through the same connection, editing files by hand where a prompt is too blunt an instrument, then calling rebuild_game to publish.

In every one of these the agent is doing the coordination, and the studio is doing the building. That division is the point. The studio's coding agent has the engine, the asset pipelines and the test harness. Your agent has the context: what you have been talking about, what you meant, what came before.

The honest limits

• Builds take minutes. A first playable lands in minutes rather than seconds, and an agent has to poll for it. The server tells agents this in its own instructions; a good one checks every minute or so and stays quiet in between.
• One account, your account. There is no way to list other people's games, read other arcades' numbers, or do anything the signed-in person could not do in the studio.
• No keys over the wire. Covered above. If an agent reports that a key is missing, the fix is in the studio.
• One export at a time per game, and not while a build turn is running, because the export needs the latest sources.
• Still invite-only. The server does not create accounts. The waitlist is the way in, and it always will be while build capacity is finite.

Try it

If you have an account, connect your agent to https://api.vbgnt.com/mcp and ask it to make something small. If you do not, leave your email and the arcade name you want, and tell us which agent you would connect first.