Coding agents
The build prompt
After provisioning, Shipyard writes a prompt for your coding agent that is specific to your stack and to what was just created: the conventions of the starter, every environment variable by name and purpose (never a value), how to send email, take payments, track events and store files, how deploys work, and a security baseline.
Print it with shipyard prompt yourproject. The same text is committed to the repo as CLAUDE.md and AGENTS.md, so an agent that opens the repo picks it up by itself.
The MCP server
Everything the CLI does is also available to an agent over MCP.
{
"mcpServers": {
"shipyard": { "command": "shipyard", "args": ["mcp"] }
}
}With Claude Code: claude mcp add shipyard -- shipyard mcp
With OpenAI Codex: codex mcp add shipyard -- shipyard mcp
The ChatGPT app only connects to MCP servers hosted on the internet, and Shipyard runs on your machine so your keys never leave it. Use Codex for that, or paste the build prompt into ChatGPT.
What an agent cannot do
- See or add a provider token. You add those in a terminal.
- Buy anything by itself. A purchase needs the amount you agreed to, passed with that call.
- Destroy a project or switch payments to live without the confirmation text from you.
- Post anything, anywhere.