Greptile: reviews and analysis

TLDR: Greptile reviews every pull request with an index of the entire repository behind it, so it sees dependencies and breakages that never show up in the diff. In exchange it talks too much, and since March 2026 it charges by volume: from $30 per seat per month with 50 reviews included, then a dollar for every extra review. It suits teams with tangled codebases and a contained change rhythm, not workflows where agents open pull requests by the dozen.

What Greptile is and how it works

Most automated code reviewers read the same material a rushed human does: the diff. They see the changed lines and a bit of the file around them, and they opine from there. Greptile starts somewhere else. Before it comments on anything it builds an index of the whole repository, with its functions, its classes and the dependencies between them, and it uses that map as the context for the review.

The difference shows up in one specific class of bug, the kind that lives not in the change but around it. You alter the shape of an object in one module and another service still consumes the old shape. You rename a field and a migration written against the old name stays behind. None of that is visible from the diff, because the diff itself is correct. It becomes visible when you have the rest of the codebase in front of you, which is exactly what Greptile is trying to have.

The second piece is custom rules. Instead of a configuration file with its own grammar, you write your house conventions in plain English, along the lines of never using a given date library or requiring every new API route to carry a permission check. The tool applies them on each review and speaks up when something drifts.

It connects to GitHub and GitLab, and that is the whole list. Anyone whose code sits on Bitbucket or Azure DevOps cannot use it today, which rules out a slice of the corporate market before price even enters the conversation.

What it is like day to day

What people who have run it report and what its documentation describes agree on the opening stretch: setup takes minutes and the first review lands on its own as soon as the next pull request opens. The initial indexing takes as long as the repository size demands, and after that comments appear before any human sits down to look.

The day-to-day problem is noise, and this is not a stray impression. The most cited third-party comparison, published on 2 June 2026 over fifty real pull requests from projects including Sentry, Cal.com and Grafana, credits it with an 82% bug catch rate against CodeRabbit’s 44%, and in the same batch eleven false positives against two. Both numbers travel together and describe the same temperament: it looks harder, finds more and gets it wrong more often too. A direct competitor, DeepSource, adds a caveat worth carrying while reading that 82%, which is that the original figure comes from the vendor’s own benchmark of fifty pull requests across five repositories rather than from an independent measurement.

The vendor has moved on this. With version 4, announced on 5 March 2026, it published that the share of comments addressed by the change author rose from 30% to 43% after a month of running both versions side by side. That is its own number, measured by itself, but at least it is dated and says where it comes from. Even so, fewer than half the comments getting addressed is still a lot of reading.

Hence the habit that decides whether a team stays or walks: narrow it down. Excluding generated directories, muting comment categories nobody cares about and writing two or three house rules turns an avalanche into a short list. Whoever installs it and leaves it as shipped ends up ignoring its comments within two weeks, which is the worst possible way to spend the budget.

Pricing and plans

There is a free plan, called Starter, with unlimited repositories, 50 credits a month and one active developer. A standard review spends one credit and a deep review spends three. The entry paid plan costs from $30 per seat per month, includes 50 credits per seat and charges a dollar for each additional credit. There is a 14-day trial, custom pricing above that, and 50% off for pre-Series A companies with under two million dollars of revenue in the past twelve months. Non-commercial MIT or Apache licensed projects that meet the requirements pay nothing.

That per-review charge is the contentious part, and it is recent. Until March 2026 the rate was flat. The change did draw public protest, though it is worth sizing before leaning on it: there is a critical page published on 30 April 2026, signed under a pseudonym, collecting cancellations and the case of open source maintainers who say they were invoiced despite being approved for free use, plus a Hacker News submission from that same day which stalled at 17 points and three comments. A real, argued complaint rather than a firestorm. The underlying claim is that the included quota was designed for a working rhythm that no longer exists, and that protest cites a developer who opened 571 pull requests in thirty days. The vendor’s answer was that fewer than 10% of active users would exceed the quota.

Before signing, do one very simple and very local piece of arithmetic: how many pull requests your team actually opens in a month, counting the ones agents open. If that number sits comfortably below the included quota, the headline price is the price. If it goes over, your real cost bears no resemblance to the website and the decision has to be made with the right figure in view.

Who it is for (and who it is not)

Greptile pays off when the codebase is large, old and tangled, with dependencies nobody holds in their head all at once. There the indexing earns back the noise: you would rather read eleven surplus warnings than lose an afternoon chasing a cross-service breakage nobody saw coming. It also suits teams with strong house conventions, because writing those in natural language is cheap in time.

It is not for anyone who measures work in pull requests per hour. With agents producing changes in chains, the billing model works against you and there are flat-rate options that do not tax throughput. Nor is it for teams whose code lives outside GitHub or GitLab, or for anyone after a quiet safety net that only speaks when it is confident, because this one speaks a great deal.

If your team sits at the other end, with few changes a month and little patience for configuration, CodeRabbit reaches usefulness sooner. The segment ranking is in the best AI code review tools and the wider map is in the best AI tools for coding.

Alternatives to Greptile

The natural comparison is CodeRabbit, which covers more platforms, charges a flat per-developer rate and comments less and with fewer false positives, at the cost of catching less according to that same comparison. If the code already lives on GitHub and review is not the priority, the pull request review bundled into GitHub Copilot comes free with what you already pay for, though it looks less deep. And for anyone working with agents, something like Claude Code can review on demand without any per-pull-request automation.

Frequently Asked Questions

How is it different from other automated reviewers?

It indexes the whole repository before it opines, rather than reading only the change. That lets it flag breakages the diff never shows, and it is also what makes it comment more than it should.

What does it really cost if my team opens a lot of pull requests?

That depends on volume, and that is the catch. From $30 per seat per month with 50 reviews included, and a dollar per review after that. Count a real month of pull requests first, then decide.

Does it work with Bitbucket or Azure DevOps?

No. Today it integrates with GitHub and GitLab only.

Is the 82% catch rate people quote reliable?

Treat it with caution. The figure comes from the vendor’s own benchmark of fifty pull requests across five repositories, so it works as an order of magnitude rather than an independent measurement.

Can I try it without paying?

Yes. The free plan gives 50 credits a month for one active developer with unlimited repositories, and the paid plans come with a 14-day trial.