Rust off Rails: The opinionated web stack for the AI era

Who Should Read

Backend developers whose coding agents pick a different stack and layout every time they build a full-stack web app

Core Mechanics

• The framework is Markdown, not code
  • It has 0 lines of framework code and 60 Markdown files.
  • The agent reads them and writes the Rust web app itself.
• The stack choices are locked in
  • Axum, Tokio, SeaORM, Askama and Tailwind are each fixed as one crate per job.
  • The agent never debates things like the ORM.
• Interview, plan, generate, green
  • It asks only product questions and writes the plan into product/ specs.
  • It is done only when fmt, clippy and tests pass on SQLite and Postgres.
• Pick one frontend mode up front
  • htmx is the default, with inertia (React/Svelte) and spa as options.
  • The api mode is still in progress.
• Eight skills and built-in dev tools
  • Skills include /offrails:new, feature, scaffold, generate, check, doctor, deploy and upgrade.
  • Traces, a profiler, a local mailbox and a jobs dashboard are built in, and the agent can read them as JSON.
• Validated by porting real apps, they say
  • The Uptime Kuma port has 1530 of 1530 tests green on SQLite and Postgres.
  • The Fizzy and FreeScout ports are in progress.

Community Reactions

• A single prompt is enough, some argued
  • One comment said "build a Rust backend using Rails-style MVC and migrations" seems sufficient.
  • Another asked how this differs from asking Claude to write best practices and then build the app.
• Docs give weaker guarantees than code, one commenter argued
  • A real framework enforces conventions in code, so this burns more tokens for fewer guarantees.
• The stack definition itself has value, one opinion
  • One commenter saw value in a "stack skill".
  • They prefer iterating in conversation and then documenting over making Markdown the source of truth.
• Similar projects already exist, commenters noted
  • Comments pointed to Loco (loco.rs) for Rust and golazy.dev for Go.
  • One commenter preferred Go over Rust for web backends because of compile times.

How to Apply

• Write a stack-lock doc for your agent
  • Today, add crates, folder layout and test rules to AGENTS.md, one line each.
  • Stop the agent from asking which library to use.
• Define done as commands
  • List lint, typecheck and test commands as the Definition of Done.
  • Make the agent report done only when all pass.
• Keep feature requests as spec files
  • Write acceptance criteria in product/features/*.md instead of chat.
  • Ask the agent to confirm each criterion maps to a test.
• Try a small app with Claude Code
  • In an empty directory, install the plugin and run /offrails:new for a small app.
  • Read the install script before piping it to a shell.

Try It

# Run in Claude Code in an empty directory (simplified version of the method)

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

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

# Next: add a feature by entering the following
Scaffold Book with title:string, author:string, published_on:date, rating:integer.
Users must be logged in to create books.

Terminology

Related Papers

• LayerFS: The filesystem for AI coding agents

  Mount an ephemeral workspace per agent, commit only the changes

• Launch HN: Alter (YC S25) – Access control for multi-user agents

  Alter gives each agent its own identity, policy, and audit trail instead of shared API keys.

• Using Claude to reduce the attack surface of code

  Claude scanned 25 open-source projects and found 493 needlessly reachable code issues.

• WhatsApp's help pages hid LLM instructions: dont recommend a regulator

  WhatsApp help pages hid 102 AI-only instructions, including not recommending a data regulator

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

Related Resources

• https://rustoffrails.org/
• https://rustoffrails.org/install.sh
• https://loco.rs/
• https://golazy.dev
• https://rustoffrails.org/docs#the-stack-in-one-breath
• https://rustoffrails.org/how-it-was-built