Hosted AI app builders — Bolt.new, v0, Lovable — are genuinely good. Until you notice three things: your prompts and your product live on someone else's server, you're paying a monthly subscription on top of raw token costs, and you're locked to their model and host.
So I built the self-hosted one. It's MIT, runs on your box with your own LLM keys, and your data never leaves.
But here's the part that's actually different. bolt.diy, Dyad, OpenHands are all a chat-to-code loop — one agent, one conversation. AICOM is a pipeline: thirteen specialized agents (analyst, PM, architect, design critic, developer, QA, security, DevOps, …) move one brief through research → design → code → QA → deploy, with quality gates between stages and an "AI Director" coordinating them. Less "AI coding assistant," more an AI dev team in a box.
One plain-language brief goes in. ~30 minutes later a real, deployed web product comes out — with automated checks that fail the build on obvious stubs or broken previews (and a hard cap of 10 repair rounds per product so a stuck build can't loop forever burning tokens).
The factory's own admin dashboard: products, a health score, pipeline metrics, and live system resources.
Honest bit — where it broke. Early on, the parallel-agent phase used asyncio.gather(...) without return_exceptions=True. So when one agent threw — a single LLM timeout — gather cancelled every other agent mid-flight and the whole run's in-progress work vanished. The fix was almost insultingly small (collect the exceptions instead of cancelling on the first one), but finding it cost a day of "why did the entire pipeline just disappear." The deeper lesson: in a multi-agent pipeline, an LLM failure is a correctness problem, not just an ops blip.
A recent addition worth a line. Producing software is half the story; the other half is letting tools get used. So there's now a small hub where you can publish a capability — a tiny HTTP endpoint — in about 15 minutes and get paid per call (aimarket publish capability.json, with a runnable examples/hello-capability and a quickstart in 20 languages), and a demand-side agent (ARGUS) that discovers and invokes them. If you've ever wanted "my API, but agents can find it and pay for it," that's the part to poke at — it's the same MIT repo.
It's all open source (MIT). Try the live demo, and if the idea resonates, a ⭐ helps a lot:
(No crypto in this post on purpose — the self-hosted/dev crowd bounces on it, and an earlier crypto-angled post of mine got exactly zero traction. The on-chain settlement story goes to Farcaster / the lottery post; here "get paid" stays a dev-tools hook.)
Top comments (0)