Machina — an opinionated, agent-native Rails template

A rails new template for building production Rails apps that are meant to be worked on by AI coding agents (Claude Code and friends) as much as by humans. It's the template ZAR uses internally — the same one our non-programmer CEO uses to ship real apps to production — with the internal plumbing stripped out so you can use it too.

"Agent-native" means the generated app doesn't just contain code — it contains the operating manual an agent needs to change that code correctly: a CLAUDE.md orientation file, 32 curated Claude skills, and 17 path-scoped coding rules that auto-apply based on which file the agent is editing.

gh repo clone obie/rails-template ~/rails-template
MACHINA_PORT=3042 rails new myapp \
  -m ~/rails-template/template.rb \
  --skip-test --skip-git

Why this exists

Most Rails templates give you a stack. This one also gives you the conventions and guardrails that keep an autonomous agent on the rails (pun intended) while it builds features for you:

• CLAUDE.md — a per-app orientation file: stack, auth flow, dev commands, conventions, and hard "never do this" rules an agent reads on the way in.
• .claude/skills/ — 32 curated skills covering Ruby/Rails craft (ruby, layered-rails, sandi-metz-rules, 37signals-style, test-driven-development), a full Hotwire pack (hwc-*), workflow skills (writing-plans, executing-plans, using-git-worktrees, handoff), and app patterns (rails-activity-timeline, rails-tiptap-autosave, mcp-oauth-setup, design-system).
• .claude/rules/ — 17 path-scoped rules (migrations, controller scoping, CRUD routing, service objects, testing conventions, schema hygiene, …). Each has a paths: frontmatter so the agent only loads the rules relevant to the file it's touching.

The result: you (or your CEO) describe a feature, and the agent builds it the way a senior Rails engineer on the team would — because the team's standards ship inside the repo.

The stack

• Rails edge (github: "rails/rails", branch: "main"), Ruby 4.0.1
• SQLite in development/test, PostgreSQL in production
• Solid Queue / Cache / Cable + Mission Control (no Redis needed)
• Hotwire (Turbo + Stimulus), Tailwind, Importmap, Propshaft
• LLM stack: Raix + OpenRouter + RubyLLM (talk to Claude / OpenAI / Gemini)
• ViewComponent, Servus (service objects), state_machines, Pagy
• RSpec + WebMock + VCR (real cassettes, no mocks)
• Sentry, Brakeman, bundler-audit, Lefthook pre-commit
• machina-auth — optional SSO layer (see Authentication)
• A production Dockerfile, a CI workflow, a justfile task runner, mise.toml for tool versions, and a wired /up health check

Deliberately not included

• Deploy infrastructure — this template is infra-agnostic. See docs/DEPLOYMENT.md for what a production setup needs and bring your own (Kamal, Fly, Render, ECS, k8s, …).
• pgvector / MCP server / media pipeline — opt-in per app; these are domain-specific and left out of the base template.

Prerequisites

• Ruby 4.0.1 — install via mise (mise use --global ruby@4.0.1)
• A recent rails gem on PATH to exec the template. The generated Gemfile pins edge Rails (github: "rails/rails"), which bundle installs afterward.
• PostgreSQL client libraries for the pg gem's C extension (brew install libpq on macOS). Dev runs on SQLite, so this only matters for bundle install and production.

Generating an app

The template is non-interactive by design — it's built to be driven by a Claude Code agent or any scripted/CI environment. Two inputs:

• MACHINA_PORT — required. The port for bin/dev and the OAuth callback. Pick a free port (3000 is almost always taken):
  hljs language-bash

  Copy

  ruby -rsocket -e 'TCPServer.new(3042).close' && echo free
  

• The app name — the standard rails new <name> argument.

MACHINA_PORT=3042 rails new myapp \
  -m ~/rails-template/template.rb \
  --skip-test \
  --skip-git

Flag notes:

• --skip-test — recommended; we use RSpec, so Minitest scaffolding is cruft.
• --skip-git — recommended; the template does its own git init + initial commit.
• Do not pass --skip-bundle — the template's after_bundle hook runs rspec:install, tailwindcss:install, db:prepare, and the initial commit.

The template aborts with a clear message if MACHINA_PORT is missing/invalid.

Quickstart after generation

cd myapp
cp .env.example .env.local      # fill in what you need (see below)
bin/setup                       # bundle install + db:prepare
bin/dev                         # web + css + worker on $MACHINA_PORT

Environment variables (all optional unless you use the feature):

MACHINA_ORG_DOMAIN=example.com   # org email domain for User#org_member?
MACHINA_PRODUCT_ID=...           # only if you keep machina-auth
MACHINA_SERVICE_TOKEN=...        # only if you keep machina-auth
MACHINA_IDENTITY_URL=...         # your identity service (SSO)
OPENROUTER_API_KEY=...           # only for the LLM stack
SENTRY_DSN=...                   # only for error tracking

Authentication

The generated app wires up machina-auth, a public gem on rubygems.org that implements SSO against a Machina Console identity service (configured via MACHINA_IDENTITY_URL). It's the standard ZAR internal auth layer, included here as a working reference — but it only talks to a compatible identity service, so for most public users it's a starting point to replace, not keep as-is.

The auth layer is intentionally isolated so it's easy to swap. To replace it:

• Delete app/controllers/auth_controller.rb, the auth routes in config/routes.rb, config/initializers/machina.rb, and the OAuth bits of app/models/user.rb.
• Remove gem "machina-auth" from the Gemfile and drop the Machina::ControllerHelpers include + before_action :authenticate! from ApplicationController.
• Add Rails 8's built-in authentication generator, Devise, or OmniAuth.

Or, if you run your own compatible identity service, just point MACHINA_IDENTITY_URL at it and keep everything.

Deployment

Infra-agnostic on purpose. docs/DEPLOYMENT.md documents the production contract (Docker image, Postgres, web + worker processes, secrets, /up health check, object storage) and how the internal ZAR version wires it up with AWS CDK, so you can reproduce the shape on any platform.

Editing the template

Two places to change things — not interchangeable:

How substitution works

Files in files/ use literal placeholder tokens that template.rb rewrites in a single pass after copying:

Most-specific tokens are replaced first (APPNAME_CLASS before bare APPNAME). There is no ERB substitution — the token approach avoids escaping headaches in JSON/YAML/shell files.

Smoke testing changes

After editing files/ or template.rb, dry-run the substitution without running the full (~10 min edge-Rails bundle) rails new:

ruby script/smoke_test.rb

It copies files/ to a tempdir, performs the rename + substitution exactly as template.rb does, checks for leftover APPNAME tokens, and parses every generated .rb and YAML file. Exits 0 on success, 1 on any failure.

License

MIT — see LICENSE. Copyright (c) 2026 ZAR / Obie Fernandez.