For AI-native developers / AGENTS.MD
AGENTS.md for a SaaS,
and what it cannot enforce
One short file that several coding agents read. useDeploy ships a real one, 40 lines, tied to the lint rules that make it more than a suggestion.
What AGENTS.md is, and who reads it
Per agents.md (https://agents.md), AGENTS.md is an open, Markdown-based file that gives coding agents project context: "a README for agents". The site lists Codex, Cursor, GitHub Copilot's coding agent, Windsurf, Gemini CLI, Jules, Zed, Warp, Devin and others as supporting it. In monorepos you can place one in each subproject; agents read the nearest file in the directory tree, so the closest one takes precedence, and explicit user chat prompts override everything. Claude Code can read it too, in place of CLAUDE.md. It is still a text file: the agent may ignore it.
What useDeploy ships today, file by file
| File in the repo | What it does | Who enforces it |
|---|---|---|
| AGENTS.md | Short architecture rules: layers, contracts, "done means green". Any agent that reads AGENTS.md gets them. | Context the agent reads. The rules it states are backed by the ESLint layer rules. |
| CLAUDE.md | Long-form operating manual: DDD layering, contracts, API regeneration, docs rules, design system. | Context only. Claude Code reads it at session start. |
| packages/eslint-config/ddd.js | ESLint layer rules: domain and application cannot import express, Prisma or BetterAuth; contexts talk through types. | Any tool, any editor. This is the part the agent cannot talk its way past. |
| .github/workflows/ci.yml | Lint and typecheck jobs run on every pull request (the test job too). | Any tool. This is the shared gate. |
| .cursor/rules/architecture.mdc | Cursor rule scoped to the server modules (layering rules, points to AGENTS.md). | Cursor only. A rule is context, not a block. |
What a real AGENTS.md contains here, and what it leaves out
useDeploy's AGENTS.md has the server layout as a four-layer table (domain, application, infrastructure, interfaces/http), seven numbered rules (dependencies point inward, framework-free core, contexts talk through types, shared contracts, regenerate the API, use the generators, "done means green") and a pointer to CLAUDE.md for the long form. It leaves out frontend design rules and anything that would not fit in one screen. It has no nested per-module files. And it states plainly that these rules are enforced by ESLint, Claude Code hooks and CI, which is only true because those files exist in the same repo.
Honest comparison: others ship agent files too
useDeploy is our product, so weigh this accordingly. MakerKit, supastarter, Nextbase, ShipAny and Indie Kit also ship an AGENTS.md, rules files or agent skills for AI tools, and AnotherWrapper sells a blocking-hook harness ("Hooks, not instructions"). What useDeploy adds on top of AGENTS.md is the combination of ESLint layer rules, Claude Code hooks and a CI gate in the same repo. It is not the only kit with agent files, and a rules file alone, ours or anyone's, is only a request.
The open-source kit and the guides
Take the guardrails without the boilerplate
The hooks and settings are published as an MIT kit: https://github.com/KonixDev/claude-code-guardrails. Drop it into any repo with an ESLint config and a guardrails script.
Write the files yourself
Guides: /blog/claude-md-file-example-saas (CLAUDE.md), /blog/claude-code-hooks-enforce-architecture (hooks), /blog/cursor-rules-nextjs-typescript (Cursor rules), /blog/claude-code-skills-subagents-architecture-review (skills and subagents).
Compare the kits
Roundup by use case, with the cases where another kit fits better.
Other AI tools
Claude Code
How Claude Code loads its rules, and what useDeploy covers for it.
Cursor
How Cursor loads its rules, and what useDeploy covers for it.
OpenAI Codex
How OpenAI Codex loads its rules, and what useDeploy covers for it.
Windsurf
How Windsurf loads its rules, and what useDeploy covers for it.
GitHub Copilot
How GitHub Copilot loads its rules, and what useDeploy covers for it.
AI coding agents: the overview
The overview of the guardrails every AI tool shares: layer rules, typed contracts and the CI gate.
FAQ
AGENTS.md or CLAUDE.md?+
Is AGENTS.md enough to keep an agent inside the architecture?+
How much does useDeploy cost?+
| Price | Founder $99 or Lifetime $199, one-time payment (no subscription) |
|---|---|
| License | Full source, commercial use, unlimited projects. Founder is a frozen snapshot; Lifetime includes every future release. |
| Stack | Next.js frontend, Express API, Bun runtime, Prisma, BetterAuth |
| Auth | BetterAuth: sessions, 2FA, magic links, password reset |
| Multi-tenancy and RBAC | Organizations with members, invitations, roles and permissions |
| Payments | Polar: full billing (checkout, customer portal, webhooks, plan sync). MercadoPago: Checkout Pro one-time payments with signed webhooks; subscriptions are not supported yet. Stripe: adapter scaffold, not charging yet. |
| Guardrails for AI agents | ESLint-enforced architecture layers, Claude Code hooks and an AGENTS.md |
| Refund | 14 days |
| Try it: free account | Sign up free and explore the dashboard: organizations, roles, billing, audit log. You buy the license from inside. https://usedeploy.dev/register |
| Docs | Documentation in English, with Spanish and Portuguese translations of the main guides |
Rules the agent reads, walls it cannot cross.
Start from a base where the architecture is enforced by lint and CI, and the agent files are already written.