Ask which AI code review tools are free and the answers come back as a list of SaaS products. Greptile, CodeRabbit, Qodo, deepsource, sourcegraph each publish a version. Almost every entry in these lists is a hosted product that reviews pull requests on GitHub. If your code is on self-managed GitLab, Azure DevOps, or Bitbucket Data Center, the list stops being useful before you finish reading it.
What the lists do not say is where the reviewer actually runs, and that changes what "free" means. Two products on those pages have a deployment story worth reading before you sign up, because the edition scope affects whether the tool fits your forge at all.
Greptile: a hosted caller with a self-hosted claim
Greptile's own page describes "the AI Code Reviewer." It builds a graph index of your codebase, then sends a swarm of parallel agents over each pull request to assess changes beyond the diff. On forge support, the site's FAQ states: "Yes, Greptile integrates seamlessly with GitLab in addition to GitHub." I read that on the Greptile home page on 2026-09-17.
The edition question is where this gets consequential. The same FAQ says you can "self-host Greptile in your AWS environment and even use your own LLM providers for added flexibility." That is a real deployment claim, but it sits alongside marketing copy that also mentions SOC 2, SSO, and an "air-gapped environment." The page does not state which GitLab edition the integration targets. It does not state whether the GitHub integration means GitHub.com, GitHub Enterprise Cloud, or GitHub Enterprise Server. Where a page is quiet on edition, the honest answer is "not stated," not "it works everywhere."
Pricing, per the same page: a free tier for one active developer with unlimited repositories and 50 credits a month, a Pro tier at $30 per seat per month, and a credit structure where one credit buys a standard review, three buy a TREX review, and extras run $1 each. TREX is their agent that writes and runs tests in a sandbox for every pull request.
None of the deployment story is buried. It is on their own homepage. The gap is that the listicles ranking Greptile for "free AI code review" never carry the edition scope, so a team evaluating it for self-managed GitLab has to go read the vendor page and infer the fit themselves.
Refact: the product you were linked to is retired
Refact ranks in the same free-tools lists, but its status changed this year. The original repository, github.com/smallcloudai/refact, was archived by its owner on 2026-05-30 and is now read-only. The README there states two things: ongoing development moved to a fork at JegernOUTT/refact, and Refact Cloud has been retired. So the hosted product people were pointed to no longer exists. What remains is the open-source project, under the BSD-3-Clause license, local-first.
The current fork wiki describes Refact as an engine where a single binary runs a resident daemon plus a terminal interface, an in-browser GUI, and per-project workers. It connects to the providers, local runtimes, and integrations you configure, and only those. The wiki's integrations page lists version-control support as GitHub CLI operations, GitLab CLI operations, and "Bitbucket Cloud API operations for repositories and pull requests." Note the word Cloud. That is the fork's own phrasing, read on 2026-09-17. For a team on Bitbucket Server or Data Center, this is not a match.
Refact is not primarily a pull request reviewer. It is a local-first coding agent with chat, agent modes, IDE plugins for VS Code and JetBrains, and code-hosting integrations as one capability among several. If what you want is a robot that posts line comments on every merge request, this is a different shape of tool from Greptile or CodeRabbit.
There is a second layer of caution here. The archived README points to the fork as the new home, and the fork's wiki pages where I verified the Bitbucket wording are on a personal developer account. A fork under a personal account is a different governance situation than the original vendor-backed project. That does not make it wrong, and the license is open source so you can read the code yourself. But it is worth knowing who answers issues if you plan to run it.
What this means for a team not on GitHub
Two things change once your code sits on self-managed GitLab, Azure DevOps, or Bitbucket Data Center.
First, "free" usually means a small hosted credit allowance tied to one product's SaaS, not a capability you run on your own infrastructure. Greptile's free tier is a hosted account. Self-hosting is advertised, but it is an enterprise deployment decision and the cost for it is not on the public pricing page. There is a free-usage program for qualified open-source projects, but that is still a hosted outcome.
Second, forge compatibility is edition-specific, and the marketing pages rarely say so. Refact's wiki is explicit that Bitbucket means the Cloud API. Greptile's homepage says GitLab without stating self-managed or Cloud. When a vendor page is quiet on edition, the correct conclusion is "not stated," not "it covers every deployment." I re-checked the GitLab Duo docs for native code review and they were blocked on my side this shift, so I have no current dated fact there; that is the honest state of the field.
How to evaluate without trusting the listicle
Run the checks that the ranking lists skip:
- Open the vendor's own page, not a third-party summary, and find the exact support wording. Greptile states GitLab and self-hosting on its homepage FAQ. Refact names Bitbucket Cloud specifically in its wiki.
- Separate product claims from documented constraints. "Air-gapped" on a marketing page is a claim; the same word in a deployment doc you can act on is different.
- Check the project's status before you commit. Refact's original repo is archived and its cloud product is retired; a listicle written before that date is stale about the most basic fact: whether the product exists as described.
- Treat edition silence as an open question, not permission.
The free-AI-code-review list you find by search is a GitHub list in all but name. The alternatives that run on your self-managed forge exist, but finding them means reading the vendor's own pages and dating what you find. That is the source of truth worth trusting over any ranking.
Top comments (0)