Coding Agents

What a coding agent working on your project can reach, what it is never allowed to do, and how to connect Stripe and Cloudflare safely.

The kit is built to be driven by a coding agent (Claude Code, Codex, Cursor) while you approve logins and paste secrets yourself. Three things keep the agent away from your secrets and your live business: rules it reads, tools that never show it a secret, and settings that block the risky outside tools.

This page is part of the Security documentation.

What Each Server Can Reach

An MCP server is a plug-in that lets the agent use a service on your behalf: the kit itself, or an outside service like Stripe or Cloudflare. .mcp.json lists the ones this project connects to. .codex/config.toml gives Codex the first two only.

ServerWhere it runsWhat it can reach
makerkitYour computer (tooling/mcp-server)This project's files, the settings in apps/web/.env (secret values always hidden as [REDACTED]), the database on this computer only, and the kit's commands. Never a database on another computer.
better-authhttps://mcp.better-auth.com/mcpBetter Auth's public documentation. Nothing of yours.
cloudflare-apihttps://mcp.cloudflare.com/mcpThe Cloudflare account you sign in with, limited to the permissions you grant when you connect it.
stripehttps://mcp.stripe.comThe Stripe account or sandbox you sign in with, limited to the permissions you grant when you connect it.

What the Agent Is Never Allowed to Do

These rules live in the project's AGENTS.md, which every agent reads first:

NeverInstead
Read apps/web/.envkit_status, kit_doctor or kit_env_read (secret values come back hidden); kit_env_update changes one setting
void secret put KEY=value, or any secret typed into a commandYou run pnpm --filter web exec void secret put KEY in your own terminal; it asks you for the value
void db execute --remote, void db studio --remoteDatabase changes as migrations (void db generate), and the deploy and go-live steps in the deployment and go-live guides
void db push on your real databaseMigrations; void db push only on a throwaway database
Ask you to paste a Stripe or Cloudflare secret into the chatYou put it in apps/web/.env or run void secret put yourself

Blocked Outside Tools

.claude/settings.json blocks the outside tools that change money or infrastructure. Claude Code refuses them even if the agent tries, and a block in the project's settings wins over anything your own settings allow.

ServerBlockedStill allowed
stripestripe_api_write (any create, update or delete), send_money, transfer_money, convert_currency, setup_recipient, atlas_company_formationReading data (stripe_api_read), searching the API and the docs, account info
cloudflare-apiexecute (it runs every request, reads and writes alike)search, docs, whoami

Changes go through the kit's own commands instead (pnpm billing:setup, pnpm kit:doctor), which show you what they will do and ask before any live or production change. Deploys run in your own terminal (Production Deploys).

The same file lets the agent run the kit's read-only commands without asking you each time: pnpm kit:doctor, pnpm kit:doctor --launch, pnpm billing:doctor, pnpm billing:setup --plan, and the tools kit_doctor, kit_status, kit_journeys, kit_prerequisites, kit_db_status, kit_env_schema and kit_env_read.

The blocks match these server names, and the official Stripe plugin for Claude Code (plugin_stripe_stripe). If you also added Stripe or Cloudflare another way (a claude.ai connector, another plugin), turn that one off: its tools have other names and aren't blocked. Codex, Cursor and other tools don't read .claude/settings.json; there, the AGENTS.md rules and the permissions you grant when you connect each server are what protect you.

Connecting Stripe and Cloudflare Safely

  • Stripe: use a sandbox, Stripe's test copy of your account (https://docs.stripe.com/sandboxes explains how to create one). When the Stripe sign-in page opens, pick that sandbox, not your live account.
  • Cloudflare: when the Cloudflare sign-in page lists permissions, tick only the ones ending in "Read". The kit's deploy runs in your own terminal with your own login.

The blocks are a guardrail, not a vault: an agent with a terminal can still run commands. What protects you is that live keys and production secrets never reach the agent, and that live and production changes need your yes in your own terminal.


Back to: Security overview