CodeRabbit on X

The good reports are specific rather than superlative. Users say it finds edge cases and logic errors that their coding model missed, and one reports a memory leak it caught during a nightly review. Several recommend it in reply to "what should I use" threads, and one describes it as an independent verification layer they could not process 200+ PRs a day without. Free reviews for open-source repos come up often, and one user says they pay for it yearly.

The pragmatic group uses it as a stage: local checks first, CodeRabbit after the push, loop until findings drop to near nothing, with Codex, Greptile or a Claude subagent alongside. One says a review loop settles once "all comes back as LGTM".

The disappointed say it cannot compete with their own repeated Opus or GPT review rounds, or call it a "slop machine". The undecided are mostly testing it — a weekend cleanup, or pairing it with Devin.

The noise complaints are about volume and shelf life. Users describe a rolling issue list that grows an unrelated item with every commit, fix loops that never close, security passes that end in thirty comments about a dummy API key in a test fixture, and notification spam. The "Fix comments" feature gets one specific knock: it fixes everything in a single session and pushes one commit, which makes the human review different.

On limits, one user calls the advertised 8 reviews an hour a "plain lie" once 72 a week trips fair usage; another says refills are calculated on a weekly average; others run out of credits on a paid plan or hit rate limits mid-iteration. The free trial's auto-renew small print catches someone.

Slowness is reported as slow or stalled reviews, a 500 error when changing plans, and one benchmark putting a review at 228 seconds against 96 for a home-built reviewer.

The trust worry is that the review bot now prompts agents to install it locally. Missed bugs appear as a user's test of six planted bugs, where they say it found 42%.

Greptile is the comparison people actually make. One user who had Fable evaluate both over 95 real PRs says each found a significant number of unique, material defects, and then that Greptile was "materially faster". Another calls Greptile more intuitive and simpler, especially for no-coders; another says it "ate CodeRabbit". Others keep both: one runs them on different PRs so whichever one has hit its rate limit hands over to the other.

Codex comes up as a replacement: a long-time free user says Codex's automatic PR review now takes over that work, and another switched to Codex to reduce spend. The in-house argument is about tailoring and cost — one user benchmarks their own reviewer at 96 seconds and $0.05 a review, with 73% of bugs found, against CodeRabbit's 228 seconds and 59%.

Single mentions go to Devin, Codacy, and a rival that drafts patches and runs tests.

The PR-count questions are sceptical rather than hostile. People ask what the 5,000 PRs were for, which features shipped in them, and "who reviewed those 5K PRs"; one says a number that size makes them question quality and intent, and another says the real test is how many reviews caught bugs that would have shipped.

The config questions are practical: which setting wins between a repo .coderabbit.yaml, a central config and org settings, and how inheritance and merging resolve; whether the tools can take context from outside the repo; whether the review modes are "terse / normal / pedantic".

On the time claim, people ask whether it holds for edge-case tests and whether it cuts time on PRs under 100 lines or mainly big ones. Smaller asks cover which bug types it finds first, whether to add it on top of a Codex × Claude cross-review, which plan was used, whether the CLI cuts rate limits, and whether a private repo must be shared for a security audit.

On findings, one person wants each one to include a concrete failing case and point to the exact diff, and wants findings marked stale automatically once the patch is revised, so an old approval does not follow new code.

On installs, one person asks for hard denies in the review agent's settings for npm i, pip install and curl | sh, arguing that a note in CONTRIBUTIONS.md is a "polite ask" that agents skip.

On secrets, one wants emails and tokens redacted from the ticket body before the agent ever reads it, or "the support inbox just moved into the model logs".

Two more are about scope: one suggests leaving GitHub so throughput stops being capped by the platform, and one says noise dropped a lot once they told the reviewer which paths are test-only and to stay quiet there unless a value leaks into a build.