Hi HN — I'm Steve. I build software mostly by directing a coding agent… — ShipDiary<br>Back to the feed<br>@customaisolutions<br>just now

Show<br>Hi HN — I'm Steve. I build software mostly by directing a coding agent (Claude Code), and the thing that kept burning me was how confident a single model is even when it's about to do something dumb on a high-stakes call — a schema change, a pricing decision, a security-sensitive edit.

So I built AI Consensus. It's an MCP server: you add it to Claude Code / Cursor / Cline, put one line in your project's CLAUDE.md ("for any high-stakes decision, consult AI Consensus first and weigh its recommendation and the strongest dissent"), and from then on your agent quietly runs those decisions past three models — Claude, GPT and Gemini — that answer independently, then cross-examine each other. You get back one recommendation plus the strongest argument against it. The point isn't forced consensus; it's surfacing where the models actually disagree, and why.

Setup is npx -y ai-consensus-connector + an API key (~2 min). Honest limitations: a run takes a few minutes (three models plus a deliberation loop), it's meant for the decisions that matter, not every prompt, and it's a paid product — but there's a free trial (2 runs, no card), and if you bring your own model keys it's unlimited and you just pay your own token cost (~$0.40–$1 a run).

I'd genuinely love feedback — especially from people who think a multi-model panel is overkill. Where does this help you, and where would it just get in your way?

ai-consensus.ai<br>10 replies

No replies yet<br>log in to reply.

Posted by<br>@customaisolutions<br>Shipped AI Consensus on day 1

Follow0 followers

About this product<br>AI ConsensusShipped

More from the feed<br>Day 1 of building ShipDiary in public.

Every builder I know has the same problem: you ship something, post about it once, and it disappears into the void. Launch platforms give you one day. Your progress updates are scattered across five platforms, and you never know which one actually brings readers.<br>So I'm building the fix, in the open: a public build diary. You write one update, it becomes tailored drafts for X, Reddit and Hacker News, posts directly to your community channels, and every link is tracked so you can see where your readers actually come from.<br>This post is entry #1 of the diary. Everything from here on gets documented: what works, what flops, real numbers included.

If you're building something, start your diary alongside mine: https://shipdiary.dev▲ 1 · 0 replies

Have a take on this?<br>Join the thread — builders reply fast here.<br>Log in to reply