Picking an AI Code Reviewer (markmarkoh.com)

A five-person engineering team trialed four AI code-review tools—Claude Code Review, Cursor’s BugBot, CodeRabbit, and Cubic.dev—as a first-pass supplement to human reviewers (not a replacement). After minimal prompt tuning they judged each on noise, usefulness, integration, and cost. Claude was the cheapest but produced single, “fluffy” PR-level comments for every update (no in-line feedback), so it was turned off. Cursor’s BugBot was the priciest (~$40/user/month) and leaned toward bug/security findings with per-line comments and a “background agent” fix flow, but came with distracting cutesy features. CodeRabbit surfaced many real issues and added diagrams/visualizations and collapsible markdown, yet generated lots of nitpicks and refactor suggestions that the team ignored; pricing: $24/user/month (annual) or $30 monthly. The team chose Cubic.dev because it struck the best signal-to-noise balance: succinct, actionable line-by-line comments, helpful PR summaries, and fewer spurious suggestions. Cubic also offers an agent-driven “Fix with Cubic” feature (untested) and uses its own review app rather than native GitHub PR reviews; pricing matched CodeRabbit. The trial highlights practical takeaways for the AI/ML community: per-line integration and concise outputs matter more than bells-and-whistles, prompt tuning can only go so far, and emerging auto-fix agents create new workflow opportunities and governance considerations.