Catch rate
100%
conclave 3-agent council, vs 100% Claude alone — both catch obvious bugs
Three frontier models — Claude, GPT‑5, and Gemini — read your pull request independently, then bring their findings to a sealed deliberation. Disagreement is reconciled in a second tier. The verdict and dissents land on the PR with the authority of all three.
Ex pluribus, iudicium. — Out of many, one judgement.
$ npm i -g @conclave-ai/cli
Open beta · 1 free hearing on installation · BYO Anthropic key for unlimited reviews · Solo $19/mo
II
proceedings
Three steps. No CI changes. No keys to manage on your end (unless you wish to bring your own).
I
One click installs the Conclave AI Code Council GitHub App. It subscribes to your repository's pull-request events; nothing runs in your CI. The work happens in our sandbox, sealed away from your build.
II
.conclave/prd.md describes what the change is supposed to accomplish — acceptance criteria, out-of-scope, non-functional requirements. Agents read it before deliberation and flag spec deviations as first-class blockers.
III
Three agents review independently. Disagreements escalate to a second tier. The verdict — APPROVE, REWORK, or REJECT — and any dissents land on the PR as a check and a comment. Solo and Pro tiers also dispatch the result to Telegram.
III
evidence
From our own dogfood — fifteen synthetic-bug pull requests across five vibe-coder Next.js templates. Honest figures; no apology, no marketing arithmetic.
Catch rate
100%
conclave 3-agent council, vs 100% Claude alone — both catch obvious bugs
Blockers per PR
10.93
vs 3.80 Claude alone — three times deeper findings, including issues lost on a single model
Spec-mismatch flags
9.0/PR
(Claude alone, with PRD attached) — blocker categories that no plain code review surfaces
How it lands on the PR
Conclave AI verdict: REWORK · 187s
4 blockers found across spec, security, a11y, and testing.
POST /billing/refund is not in the PRD's acceptance criteria. Either remove the route or extend .conclave/prd.md §3 to include it; shipping refund flow without spec coverage is exactly the audit-trail gap legal flagged last quarter.apps/api/src/routes/billing.ts:84dangerouslySetInnerHTML renders the user-supplied bio field unchanged. XSS-trivial. Sanitize via DOMPurify or use plain text + line breaks.apps/web/src/components/Profile.tsx:31aria-label. Screen readers announce nothing — the modal becomes inescapable for keyboard / VoiceOver users.apps/web/src/components/Modal.tsx:18/billing/refund route ships with zero integration tests. Refund logic untested in the same PR that introduces it inverts the safe order — add a test before merge.apps/api/test/billing.test.ts (missing)Push a fix or run conclave autofix --use-saas --pr 142 to let the worker agent attempt the fixes automatically.
Verbatim format the council posts after a verdict. Real categories, real chips, real autofix command — no marketing reshoot.
The moat is not “smarter than Claude.” It is multi-agent depth and PRD-aware spec compliance — three models reading your change and your spec together catch scope creep, route mismatches, and forgotten acceptance criteria a single model alone misses.
IV
hearing
Paste a public pull request URL or a raw diff. Optionally attach a PRD by drag-drop, file picker, or a GitHub URL. One Claude pass with the PRD-aware prompt — the same prompt the full council uses. Three opening hearings per IP per UTC day; install the App for the full three-agent council.
V
indulgences
Hard cutoffs, no surprise invoices. Bring your own Anthropic key for free unlimited usage, or pay $3 once for a single council pass on one PR.
Free
$0bring your own Anthropic key
First-PR pass
$3once · no subscription
Solo
$19coming soon
Pro
$49coming soon
Free + First-PR pass are live today. Solo and Pro subscriptions land once we have paying users — DM @baessi1 on Threads to join the waitlist. Lemon Squeezy (Merchant of Record) handles VAT across KR / US / EU so we don't have to.
VI
questions