Hey Product Hunt 👋 I built CodeSolar after watching a team quietly switch off yet another AI reviewer. The pattern is always the same — exciting for a week, then muted. Not because AI can't read code, but because of three specific failures: 1. Comments on the wrong line. The most common way LLM review breaks. We render real line numbers into the diff we hand the model, then re-verify every number it returns. If a finding's evidence lines aren't in the diff at all, we drop it instead of posting it. A finding the model marks "probable" can't wear a red label. 2. It re-reviews the same code forever. Push a commit, get the same three comments back. CodeSolar carries the previous review forward: follow-ups read only the new commits and report what you fixed ("2 of 3 resolved"). Dedupe isn't left to the prompt — near-identical findings are filtered before posting, even after lines shift. 3. It doesn't know your team's rules. When someone with write access replies "this empty catch is intentional," that reply becomes one lesson applied to later reviews of that repo. Every lesson is visible in settings with a link to the original thread, and one click switches it off. Lessons can only change what gets looked at — never the review rules themselves. That's a prompt-injection surface and we treat it like one. Powered by Upstage Solar-Pro4. Reviews in English · 한국어 · 日本語, it reads your AGENTS.md or CLAUDE.md so it knows your conventions, and setup is: sign in with GitHub, pick repos. No CI to wire, no runners, no YAML. Free during beta. I'd genuinely love to hear where it gets things wrong — that's the feedback I'm here for. 🔆

CodeSolar

AI code review for GitHub that never nags you twice

AI code review for GitHub that never nags you twice