# Build with AI

PayLynks is built to be wired up by AI coding tools (Claude Code, Cursor, Lovable, Bolt, Replit) on the first try, safely.

| Piece | What it gives the AI |
| --- | --- |
| [llms.txt](/llms.txt) and every doc as `.md` | Current docs instead of stale training data. Append `.md` to any docs URL. `/llms-full.txt` has everything in one file. |
| [MCP server](/docs/mcp) at `https://paylynks.io/api/mcp` | Tools to create links, invoices, customers and payment intents, look up payments, refund in test mode, fetch working code. |
| [CLI](/docs/cli): `npx paylynks init` | One command from a blank app to a working test checkout, with keys written to `.env.local`. |
| [Snippets](/docs/checkout) | Copy-paste code for Next.js, React/Vite, Supabase, plain HTML and Express. |
| [Agent skill](/docs/agent-skill) (`/SKILL.md`) | The do's and don'ts: where keys go, how prices are set, test cards, common errors. |

## Tell your AI tool

> Add payments with PayLynks. Read https://paylynks.io/llms.txt first, then run `npx paylynks init`.

## Guardrails

- **Agents never see secret keys.** MCP connections use scoped agent tokens. The CLI writes keys to `.env.local` and nowhere else.
- **Read-only by default.** Creating things (`write`) and refunds (`refunds`) are separate permissions you choose when connecting.
- **Spend limit.** Each connection has a per-call limit; larger links, invoices or refunds are refused.
- **Test mode.** Agent connections are test mode. Live refunds always need a person in the dashboard.
- **Logged.** Every tool call is recorded with the client's name (event `agent.tool_called`).