Rust off Rails 发布面向 AI 时代的 Rust Web 框架

A web framework that ships zero lines of code

Rust off Rails is a set of Markdown files that tell coding agents how to build a full-stack Rust web app: one blessed stack, one layout, one way to do each thing. The convention lives in the prose. The agent writes the code.

curl -fsSL https://rustoffrails.org/install.sh | sh

or in Claude Code: /plugin marketplace add rust-off-rails/rustoffrails

Read before you pipe: the install script · INSTALL.md

• 0 lines of framework code
• 60 Markdown files
• 2 databases, same code

What it is

A framework made of decisions, not code.

Other frameworks ship a library. Rust off Rails gives your agent the judgment: every choice it would otherwise improvise, already made and written down.

The founding rule

All code is written by AI

Every line of Rust, SQL, HTML, CSS, TypeScript and every Dockerfile is written by an agent. Humans write intent: specs in product/, answers in the interview, feedback in conversation. Never code. It's the rule the others rest on: boilerplate is free, conventions are enforced by the agent, and specs stay the source of truth. Read §0.

Markdown is the framework

Where a framework would ship a library, Off Rails ships a precise spec. The job queue, sessions and Inertia adapter are rendered into your app as code you own.

Strong opinions, aimed at agents

Crate choice, file layout, naming, errors, tests: decided once. The agent never debates the ORM. Inconsistency is the enemy, not imperfection.

Batteries included

Auth, jobs, email, caching, uploads, realtime, deploys: every layer has a guide and a recipe, so the agent never starts from a blank page. Browse the guides.

How it works

Interview. Plan. Generate. Green.

The agent asks only what a human can answer, writes the answers down, and is done only when the compiler and the test suite on both databases agree.

1. 01

   Interview

   Product questions only, each with a recommended default. "Who can see this?" yes; "which ORM?" never. Say "you decide" and it takes the defaults.

2. 02

   Plan

   A compact plan (resources, pages, jobs, assumptions) written into product/ specs with acceptance criteria. The spec is the durable record; chat is not.

3. 03

   Generate

   Generator recipes, run by the agent: migration → model → failing tests → controller → views, one vertical slice at a time.

4. 04

   Green

   fmt, clippy, tests on SQLite and Postgres, every acceptance criterion traced to a test, coverage gates. Green or it didn't happen.

Eight skills, one per job

How you'll use it

You write intent. The agent writes everything else.

1. 1

   Say what you want, or write it in product/features/*.md. Plain English.

2. 2

   The agent writes the spec, then the code: migrations, models, views, tests.

3. 3

   Done means green on SQLite and Postgres, every acceptance criterion tested.

Read the guide

Dev tools

See inside your app.

Traces, a profiler, a local mailbox, a jobs dashboard and live presence, built in. Your agent reads the same data as JSON, so it can check its own work.

• Dev dashboard

• Traces

• Hot functions

• Mailbox

• Jobs

• Presence

The dashboard costs about 1 µs per request in a release build, and is development only by default.

Take the tour

The deal

What changes when typing is free.

A library with hundreds of thousands of lines

60 Markdown files

A code generator binary

generators/scaffold.md: a recipe the agent runs

Convention over configuration

Convention over deliberation: the agent never debates which crate to use

Optimized for programmer happiness

Optimized for agent correctness: compiler-checked everything

Magic, to save keystrokes

No magic. Explicit code is cheap when you don't type it.

The stack in one breath

One crate per job. Substitutions are not taken at the table.

Axum·Tokio·SeaORM SQLite locally, Postgres in production, same code·Askama·Tailwind·lettre·tracing·one binary, one Docker image

FRONTEND_MODE, chosen once in the interview

htmx default

Askama templates, checked at compile time, plus htmx. No Node. For CRUD apps, dashboards, content sites (like this one).

inertia

React or Svelte pages with server-side routing and controllers, through a generated Inertia adapter. The pick when you want a JS UI.

spa

A separate single-page app plus a JSON API, for when the UI must be a standalone app. Types flow from Rust via ts-rs.

api in progress

No UI at all: a backend and a versioned API for mobile apps, partners and scripts.

UI_LIBRARY

react (default) or svelte, for inertia, spa and htmx islands

API

none, rest (OpenAPI + tokens), graphql or both: a public, versioned API, independent of the UI

And in the database: a job queue, a cache and pub/sub. The database is the only infrastructure.

Standing on giants

The best of Rails and Phoenix, together.

None of the good ideas here are new. We took the ones that aged best from two frameworks we love, gave each a Rust shape, and wrote them down for agents.

From Ruby on Rails

• Convention over configuration, sharpened into convention over deliberation
• Generators: rails new and scaffold, as recipes the agent runs
• Active Record → models and migrations
• Rails 8 authentication: database-backed sessions, generated and owned
• Solid Queue, Cache and Cable: the database is the infrastructure, no Redis
• Turbo 8 page-refresh broadcasts → signals over SSE
• Action Mailer previews
• Kamal, /up and Docker by default
• bin/rails tasks → one app binary
• Credentials and env discipline

From Elixir and Phoenix

• Oban: persistent jobs with states, unique jobs, cron, snooze
• Oban Web → the live /admin/jobs dashboard
• Swoosh's local mailbox → /dev/mailbox
• phx.gen.auth: magic links and the "check your mailbox" dev hint
• LiveDashboard → the /dev toolbox
• The pending-migrations page, with its "Run migrations" button
• Ecto's explicit data layer: changesets, no callbacks
• Channels, Presence and PubSub in progress
• Distributed nodes and "let it crash" supervision in progress

And also

• htmx
• Inertia.js
• Tailwind
• shadcn-style owned components
• ts-rs end-to-end types
• Svelte
• Playwright

Field tests

Real apps, ported by agents.

Agents port well-loved open-source apps to Rust off Rails. Every rough edge they hit goes back into the framework.

• from Fizzy In progress

  fizzyer

  Fizzy's boards and cards, rebuilt in Rust by agents.

• from FreeScout In progress

  A FreeScout port

  A shared-inbox help desk, built by agents on another machine.

• from Uptime Kuma Built

  uptime-offrails

  1530 of 1530 tests green on SQLite and Postgres.

Built with GPT-6 Luna Claude Sonnet 5.5 Meta's Muse Wonder 1 GLM 5.3 Flash via opencode The goal: small, cheap agents that get real work done too.

Read the field tests

Get started

Install the plugin. Describe the app. Watch.

1 With Claude Code (recommended)

/plugin marketplace add rust-off-rails/rustoffrails
/plugin install offrails@offrails

Then, in an empty directory:

/offrails:new bookshelf: people track the books they've read and review them

The agent interviews you, confirms a plan, then builds. Skills also trigger on plain requests ("make me a new Rust web app…").

2 With Codex, Cursor, or any agent that reads AGENTS.md

Read /path/to/rustoffrails/AGENTS.md and make me a new app called "bookshelf".

Generated apps carry AGENTS.md and a vendored .offrails/, so later sessions need no setup.

3 Then

Describe features in plain language, or write them as specs in product/:

Scaffold Book with title:string, author:string, published_on:date, rating:integer.
Users must be logged in to create books.

The full walkthrough How you'll use it Read the docs Star on GitHub