I pointed a "team of AI agents" at a real feature request. It actually compiled.

A developer tested ohmyaistaffs, a self-hosted, single-binary AI tool that claims to turn plain-language requests into compiled, tested applications. The tool uses a team of AI agents (analyst, designer, engineer) and validates output with Go build, vet, and test before completion. The developer found it surprisingly effective for generating a working prototype, though it is early-stage with future paid-tier scaffolding.

I'll be honest, I clicked into this one ready to be annoyed. Every other week there's a new "AI dev tool" that turns out to be a chat window with a system prompt and a GitHub star-begging README. I've burned enough evenings on those to be suspicious by default. But a friend pointed me at ohmyaistaffs https://ohmyaistaffs.io last week — a self-hosted, single-binary thing that claims to take a plain-language request and hand you back a compiled, tested, running app , not just a pile of markdown or a scaffold with 40 TODOs in it. No account, no license key, free. That combination of claims self-hosted + no login + "it actually compiles" was specific enough that I figured it was worth ten minutes to check if it was real. No signup flow, no dashboard to click through. Just: curl -fsSL https://ohmyaistaffs.io/install.sh | bash That's it — it drops a single Go binary on your machine and serves its own UI locally. No Docker Compose file, no Postgres you have to spin up first, no "create an account to get your API key." If you already have Claude Code CLI installed and logged in, it'll piggyback on that session instead of asking you for a key at all. If not, it supports basically everything else too — Anthropic, OpenAI, Azure OpenAI, Bedrock, Vertex, Gemini, OpenRouter, Groq, DeepSeek, Mistral, xAI, or fully local via Ollama/LM Studio. Bring your own model, swap it later, nothing's locked to one vendor. This is the part I expected to be underwhelming. It wasn't quite. You type one sentence — I tried something boring on purpose, a small internal tool for tracking equipment loans — and instead of immediately vomiting out code, it comes back with clarifying questions first. Who can mark something as returned? Does overdue matter? Is there more than one location? It's a small thing, but it's the difference between "the model guessed and baked a wrong assumption into 12 files" and actually catching that stuff before code exists. After that it works in three visible stages, and you can stop at any of them: The UI represents this as agents sitting at desks on a little office floor with a task board queued / working / in review / done , which sounds gimmicky written out like this, but in practice it's just a legible way to see which of the three roles — analyst, designer, engineer — is doing what, and there's a clear "waiting on you" indicator when an agent is stuck on a decision instead of silently guessing. Anyone can promise "AI writes your app." The thing that made me trust this more than the average tool: the generated backend is plain Go stdlib, and the pipeline runs go build , go vet , gofmt , and go test against what it produces before calling a task done. It's not "trust me, the LLM said it works." There's a compiler in the loop. Storage is JSON files with atomic rename writes rather than some hidden managed DB you now depend on, and CRUD operations do real state-transition validation instead of the classic AI-demo move of stubbing everything as a 501. Two other details worth mentioning if you're evaluating this seriously rather than just kicking the tires: package.json scripts, go.mod , a Makefile target, whatever — and gives you a live preview against your actual codebase.It's not magic, and it's not going to replace someone thinking about your actual product. It's also clearly early — there's dormant licensing/billing scaffolding in the codebase for a future paid tier that isn't wired up to anything yet, which tells you it's a project still figuring out its shape, not a finished commercial product with a pricing page. If you need enterprise SSO or a hosted managed version today, this isn't that — it's a binary you run yourself. But for the specific job of "turn a vague internal request into a working prototype I can poke at before committing to a real design doc," it did what it said, and the fact that the output has to survive go build before it's called done is the reason I'd actually reach for this again instead of just letting it collect dust next to the other tools I tried once. If you want to try it yourself, it's still just: curl -fsSL https://ohmyaistaffs.io/install.sh | bash Curious if anyone else has run it against something bigger than a toy internal tool — drop a comment if you have, I'd like to know where it breaks.