Blog · 16 November 2026 · 6 minute read
An MCP server for launching apps from Claude Code, Cursor and Codex
Your AI coding tool can write the app. With an MCP server it can also set up the domain, hosting, database, email and payments. How that works, and how to keep it safe.
What MCP adds to a coding tool
The Model Context Protocol, MCP for short, is a standard way to give an AI tool new abilities. An MCP server offers a list of tools, each with a name and a description, and the AI calls them when they help with what you asked. Claude Code, Cursor and Codex all support it.
Without MCP, your coding tool can write files and run commands. With the right server it can also act on the world outside the repository in a controlled way.
Why launching is a good fit
Launching a product is a long list of small, exact steps at several providers. An AI is good at following a list and bad at guessing what exists. A launch server gives it both: tools that perform each step, and answers that describe what was created.
So instead of asking the AI to write a guide you then follow by hand, you ask it to launch the product and it calls the tools.
What a launch server should expose
- Read-only tools to list connected accounts, show the chosen stack and report what a launch would do.
- A rehearsal that walks through every step without creating anything.
- The launch itself, with progress the AI can report back.
- A status tool to see what exists for a project.
- A removal tool that lists what it would delete and needs explicit confirmation.
The safety questions to ask
Giving an AI the ability to create infrastructure deserves care. Ask these of any server before you connect it.
- Where do the provider tokens live? They should stay on your computer, ideally in the system keychain, and never be shown to the model.
- Can it spend money without you? Purchases should show a price and wait for a person.
- Is there a spend limit per launch?
- Does it mark which tools only read and which can change or delete things?
- Does it redact secrets in everything it returns?
- Does it run locally, or do your tokens pass through someone else's server?
How Shipyard does it
Shipyard runs on your own computer and includes an MCP server next to its command line. Tokens are stored in the keychain and are not sent to our servers or handed to the model. Each tool is labelled as read-only or as one that can change things, so the coding tool can ask you before the second kind.
A purchase, such as a domain, is shown with its price and needs your confirmation. A spend cap applies to every launch. Removing a project lists exactly what will go and asks for the project name, and a domain registration is never deleted.
What it looks like in practice
You open your coding tool in an empty folder and say what you want to build and what it should be called. The AI checks that the domain is free, shows the stack and the price, and asks you to confirm. It then runs the launch and reports each step.
When the launch is done, the AI receives a build prompt that describes the live setup: which host, which database, which variables hold which keys. It writes the application against that, pushes, and the host deploys it to your domain.
Limits worth knowing
- An MCP server that runs locally works in tools on your computer, not inside a chat on a phone.
- You still create the provider accounts and tokens yourself, once.
- The AI can call the tools, and you remain the one who confirms purchases and removals.
Launch in an hour, not a weekend
Shipyard connects your accounts once, then sets up the domain, hosting, database, email and payments for every new product.
See how it works