Manki sizes each review to the nature of the change: one reader for a change where no defect could hide, two readers on different models for one that could reach your users. Dedup drops repeats from prior reviews, a judge filters noise and classifies what remains, and manki learns your team’s conventions over time.

Why Manki

• Multi-stage pipeline — a planner triages the change, finders review in parallel, dedup catches repeats from prior reviews, a judge filters noise and classifies findings by severity
• Two models where it matters — changes that could reach a user get a second finder on another model that searches from the code around the change, so the two readers miss different things
• Smart verdicts — blocks PRs on real issues, approves over style nitpicks; auto-approves when all blocking threads resolve
• Multi-provider — Anthropic, OpenAI, and Gemini, mix-and-match per stage via .manki.yml
• Self-learning memory — /manki remember teaches conventions, /manki dismiss suppresses false positives, patterns stick across PRs
• Conversational — reply to any review comment to discuss, or use @manki explain/forget inline
• **Self-hosted GitHub…