How to use a starter kit

1. Download the kit #

On a recipe page, press Download the kit (.zip). That .zip holds the files for Claude Code with the default answers. For another agent or other limits, press Change parts or your agent: the kit builder opens with the same recipe, and its download matches your choices.

2. Open your agent in an empty folder #

Unzip the kit into a new, empty folder and keep the files at its top level. Then start your agent there.

Run claude in the folder. Claude Code reads CLAUDE.md, which holds one line that imports AGENTS.md. It holds the MCP servers of the project's .mcp.json as "Pending approval" until you accept them, so approve them when it asks. The kit builder also shows claude mcp add commands that add the same servers from the terminal.

Pick Cursor in the kit builder before you download. Open the folder in Cursor: it reads AGENTS.md at the project root, and the kit's .cursor/mcp.json lists the MCP servers for Cursor.

Pick OpenAI Codex in the kit builder. AGENTS.md is the native instruction file of Codex. The MCP servers are not a file in the kit: the builder shows [mcp_servers.…] tables to add to your own ~/.codex/config.toml (a project can use .codex/config.toml once it is trusted). Then run Codex in the folder.

Pick GitHub Copilot (VS Code) in the kit builder. Open the folder in VS Code. Copilot reads AGENTS.md when the chat.useAgentsMdFile setting is on, and the kit's .vscode/mcp.json lists the MCP servers.

Pick Another agent in the kit builder. You get mcp.json in the common mcpServers format: move it to where your agent expects its MCP config. Most agents read AGENTS.md at the project root; Gemini CLI needs it added to context.fileName, and Aider needs --read AGENTS.md.

3. Paste the first prompt #

Paste the text of first-prompt.txt as your first message. By default it asks the agent to read AGENTS.md, do step 1 of the plan and stop, then report the commands it ran, what each check showed and anything in the brief that was wrong. You reply, and it does the next step.

If you would rather let it run, the recipe page also has a whole-plan prompt: the agent works through every step, checks each one itself, writes a line per step in PROGRESS.md and stops when the definition of done is met.

4. Follow the plan #

AGENTS.md splits the build into steps. Each step ends with Done when lines, and the agent does not start a step until the previous one meets them. The last step checks the whole Definition of done. The brief tells the agent to stop and ask you for an account, a key, a payment, or any part it does not list.

5. What you check yourself #

Left to you: the recipe page lists what no check can do, such as creating an account or choosing where the project runs.
Keys and accounts: you create them and put them where the brief says, usually a .env file. The agent stops and asks instead of signing up for you.
Licenses: CREDITS.md records every third-party file, and the agent adds a row for each file it brings in.
Draft kits: read Check these yourself at the bottom of the kit before you start. See Tested kits and draft kits.

The files in a kit #

File	What it is
`AGENTS.md`	The brief: what to build, stack and versions, licenses, plan with checks, known traps.
`first-prompt.txt`	Your first message to the agent.
`install.txt`	Install commands with exact versions.
`CLAUDE.md`	One line that imports `AGENTS.md` for Claude Code (Claude Code kits only).
`.mcp.json`	The MCP servers the brief lists, for Claude Code. Cursor gets `.cursor/mcp.json`, Copilot `.vscode/mcp.json`.
`CREDITS.md`	The license record of every third-party file.
`kit.json`	The kit as data: parts, versions, checks.

If something goes wrong on the way, see When the agent gets stuck.