Rust off Rails ships 60 Markdown files that agents build from

A framework used to compete on the lines of code it saved you from writing. Rust off Rails launched this week with the opposite bet: that if an AI agent is writing every line, the point is not to ship a library but to ship the decisions. The Hacker News discussion of its launch picked up 75 points on 21 October.

The project is a set of 60 Markdown files that tell an AI coding agent how to build a full-stack Rust web application: one approved stack, one layout, and one way to do each job. The convention lives in prose, and the agent writes the code. The repository is MIT OR Apache-2.0 licensed and the website lists its current release as v0.7.3.

What is in the box

The stack is picked once and the agent does not debate it: Axum and Tokio for the server, SeaORM for the database, Askama for templates, Tailwind, lettre for mail, and tracing for logs. The database runs on SQLite in development and Postgres in production from the same code, and holds the job queue, cache, and pub/sub, so Redis is not a dependency.

The agent is given eight skills, each a short Markdown recipe, including /offrails:new for a new application, /offrails:feature for one spec-first change, /offrails:scaffold for a resource end to end, and /offrails:check, which runs cargo clippy, fmt, and the test suite on both SQLite and Postgres before anything is called done.

The interview

A new project starts with a product interview the agent runs in plain language. Rust off Rails says the questions are the ones only a human can answer: accounts (magic link, email and password, invite-only, or none), who owns data, the frontend mode (htmx, Inertia with React or Svelte, a separate single-page app, or an API-only backend), and whether to expose a public REST or GraphQL API. The answers are written into the project's own product/ directory as the durable spec, and the agent then generates migrations, models, failing tests, controllers, and views in vertical slices until the test suite is green on both databases.

Field tests

Rust off Rails lists four apps its team has asked agents to port as field tests: Fizzy's boards, FreeScout's shared inbox, Mastership's Uptime Kuma, and the project's own site. The site's claimed port, uptime-offrails, has 1,530 of 1,530 tests green on SQLite and Postgres per the project's own page, with agents named as GPT-6 Luna, Claude Sonnet 5.5, Meta's Muse Wonder 1, and GLM 5.3 Flash via opencode. These are the project's own figures.

How a team adopts it

In Claude Code, the install is two commands at the plugin marketplace and then /offrails:new <description> in an empty directory. In Codex, Cursor, or any agent that reads AGENTS.md, the agent is told to read the Rust off Rails AGENTS.md file and build the app, and the generated project vendors .offrails/ so later sessions need no setup.

Who changes what tomorrow

For a Rust team that already runs an agent on each change, the trade is clear: give up the freedom to pick a different ORM on every project in return for a convention the agent will not re-litigate. Rust off Rails states the cost of that trade in its own words: substitutions "are not taken at the table", and swapping Axum or SeaORM out means forking the framework. The right test is to run /offrails:new on a small internal tool and compare the output to the application that team would have written by hand.

Source

Rust off Rails and its Hacker News launch thread.