Hey PH 👋 I'm Seunghun, solo on this. I ship almost everything with agents now, and I kept rubber-stamping PRs that *ran fine* but weren't what I'd asked for — an extra field nobody wanted, a requirement quietly dropped, an auth check that drifted from the server into the UI. Nothing in the diff was wrong, so my single-model reviewer had nothing to say. It only looked wrong next to the spec. So Conclave runs three models as a council on each PR (they actually disagree, then I get one verdict), and if you attach a PRD it flags that scope drift directly. There's an autofix worker, and it learns from what I merge vs reject. I dogfooded 15 synthetic-bug PRs across 5 Next.js templates: same catch rate as one model, but ~3× the blockers surfaced (10.93 vs 3.80/PR). The extras were boring-but-real — missing tests, edge cases, a string-concat query. (It's an n=15 internal run, indicative, protocol's in /benchmarks — please try to break it.) It's not a Cursor/Copilot replacement. It sits at the PR layer as a merge gate. Free on public repos, self-hostable (FSL). Two things I genuinely want PH's read on: 1. Does "review against the spec" matter to you, or do specs go stale too fast? 2. With Anthropic's own Claude Code Review now live, is a 3-model council differentiated enough? I'm here all day — point me at a public PR and I'll post what it catches.

Conclave AI

A code-review council for AI-written PRs

A code-review council for AI-written PRs