Remy is an AI agent that builds full-stack TypeScript apps from a spec written in a new flavor of annotated markdown. The spec has two layers: prose describing what the app does, and annotations that carry the technical precision (data types, edge cases, validation rules, code snippets). The agent then "compiles" this into code: backend methods, typed schemas, frontends, test scenarios, and everything else are derived artifacts of the spec.
The idea is that this isn't no-code, and it isn't a shortcut for people who can't code. It's a step toward a new kind of higher-level programming language, one that happens to look like annotated prose instead of syntax with semicolons. The spec isn't a requirements document that generates code. It is the program. Code is compiled output, the same way nobody writes assembly by hand anymore but it's still running underneath. The spec and code stay in sync bidirectionally, and the same spec compiles into different interfaces (web app, REST API, conversational AI agent, MCP server, etc.) depending on what you need. We're obviously stretching the word "compile" here - the same spec isn't going to produce character-identical code each run. But we think it's an interesting mental model for thinking about where software is headed, and we think the gap between "functionally equivalent" and "effectively deterministic" is only going to continue narrowing with each generation of model.
At their core, the apps Remy builds are git repos and markdown files, but we've also built a browser-based sandbox around them - editor, live preview, terminal, chat, and all the other bells and whistles - with managed infrastructure underneath: SQLite with automatic schema migrations, auth primitives (sms/email verification codes, sessions - you build your own UI and define your own user table), 200+ AI models and 1000+ integrations as SDK calls, deploy-on-push. Standard TypeScript, any npm package, any frontend framework.
It's been really fun to see what is possible with all of these pieces connected. We're opening up in public alpha this morning. You can watch some demo videos, see some more details, and sign up to try it at: https://remy.msagent.ai (sign up for a free MindStudio account and then use "showhn" as the code to join the alpha without needing a paid account)
There are also some more demo videos on YouTube: https://www.youtube.com/watch?v=2QJvBcAqQqA&list=PL1gRZlpf9_...
As well as a tour of the interface: https://www.youtube.com/watch?v=ar724yPXgwU
Would love to know what you think. Feel free to drop me an email at sean at mindstudio.ai too
elllyphant•24m ago
The problem stems from nobody sharing a good document of "what this is what the app is supposed to do." The codebase became the only record of intent, and it was written by AI.
Spec-driven development fixes that. You can hand it to a new teammate, read & understand it, and actually understand the system. The agent reads the same document and generates from it. When something needs to change, you change the spec, not a pile of files you didn't write that become messy later.
We're really early and want feedback from people actually building things. If you try it, please share your experience!