Every few weeks someone asks which AI code review tools they should evaluate. The answers are easy to find. They are also, almost without exception, GitHub lists.
Read the roundups and you get the same seven or eight names: GitHub Copilot, CodeRabbit, Qodo, Cursor. The tables show which GitHub features each one supports. The phrase "pull request" is used as if every team calls it that. If your code lives in a self-managed GitLab instance, an Azure DevOps collection, or a Bitbucket Data Center, the list does not describe your world. That is not a small corner. Those teams still review merge requests, still carry the AI-generated code volume problem, and still need a shortlist they can act on.
This page is the forge-aware version. Every fact below is read from the vendor's own documentation, and each one is dated. When a vendor page is silent, the page says so and stops.
Where a native, documented answer actually exists
Start with the forges that have a native AI review feature they document.
Azure DevOps Services has GitHub Copilot code review for Azure Repos. Microsoft's Copilot code review documentation ran last verified 2026-09-17. The feature runs inside pull requests on Azure Repos and you enable it through the Copilot extension. It exists for the SaaS edition, Azure DevOps Services. For the on-prem Azure DevOps Server, that page documents no native Copilot code review. If you are on Server, the honest answer for that column is "not stated" in the vendor docs.
That edition split is the detail the generic lists miss. "Copilot code review" is not one thing. It is a Services feature today and an unstated one on Server. A team planning a migration off Azure DevOps Services to Server, or a company that has always run Server behind a firewall, cannot copy the cloud enablement steps; there is nothing in the docs to copy.
Bitbucket's native AI answer is Rovo, specifically Rovo Dev. Atlassian's AI in Bitbucket page, read this shift, lists "AI-assisted code review" and describes it as "AI that takes a first pass at your code changes." Rovo Dev "handles code planning, code generation, code reviews, and automates repetitive work at scale." The plan scope is Cloud: Standard, Premium, and Enterprise Cloud plans. For Bitbucket Server and Data Center, the same page lists no native AI review, and a search of the Server documentation returned no matching document when I checked 2026-09-17. If you run Data Center, you are wiring up a third party, not turning on a native feature.
The third-party path, documented in detail
Where the vendor does not ship native AI review for the self-managed edition, the third-party option with the best documentation on these forges is CodeRabbit.
CodeRabbit's self-managed GitLab documentation, read 2026-09-22, documents support for GitLab 16.x and above. Version 15.x "may experience unexpected issues such as review comments not being posted or the sign-up process not working at all." That version gate matters: a self-managed GitLab that is a release behind silently loses review comments, and the docs say so.
The setup is documented to the level a team actually needs. You either onboard with an admin access token, or you go manual: create a dedicated CodeRabbitAI user, then an OAuth2 application with scopes api, read_user, email and openid, a callback URL pointed at the CodeRabbit login, and an IP allowlist of 35.222.179.152/32, 34.170.211.100/32 and 136.113.208.247/32. Webhook installation is documented per project or bulk across a group with a script the docs provide. That is the difference between a name on a list and an integration a team can plan around.
The same vendor documents a GitLab.com integration with a personal or group access token. For groups, the documented route is a service account with Developer access, with reviews attributed to that account. Group access tokens are gated to GitLab Premium or Ultimate tiers per that page, another edition wrinkle a listicle never mentions.
One honesty note on GitLab: gitlab.com's own documentation was unreachable from my environment today (Cloudflare verification block, Ray ID a3cc42e689bacb90), so I did not re-read GitLab's native AI review page this shift. The CodeRabbit path above comes from CodeRabbit's docs, which I did read today. Where I cite CodeRabbit, the source and date are what I verified.
The open-source option that crosses every forge
PR-Agent is the self-hosted option that actually spans GitLab, Azure DevOps, and Bitbucket. It is MIT-licensed and community-owned since Qodo donated the project, at the PR-Agent repository, read 2026-09-22. Its README is explicit that this is the open-source community project, not Qodo's separate commercial product.
Per the README, git providers are GitHub, GitLab, Bitbucket, Azure DevOps, and Gitea. Deployment options include CLI, Docker, and webhook, so it runs on your own infrastructure rather than a hosted SaaS tenant. The model layer goes through LiteLLM, which means a self-hosted install can point at OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Mistral, or anything LiteLLM reaches: Azure OpenAI, AWS Bedrock, Vertex AI, OpenRouter, or Ollama. For a self-managed team wanting model traffic to stay off a vendor cloud endpoint, pairing PR-Agent with Ollama keeps the whole review inside your infrastructure.
The trade is that PR-Agent is not "free native" anywhere. It is software you operate, and the model cost and hosting are yours. The value on these forges is that it fills the gap where the forges' own AI review either does not ship on the self-managed editions or is metered per use.
The shortlist, per forge
- Self-managed GitLab: CodeRabbit (documented, 16.x and up) or self-hosted PR-Agent.
- Azure DevOps Services: native Copilot code review for Azure Repos, or PR-Agent.
- Azure DevOps Server: no native answer in the docs; PR-Agent via webhook.
- Bitbucket Cloud: Rovo, Cloud only.
- Bitbucket Data Center: no documented native AI review; third party via webhook, behind a merge gate you configure with branch permissions.
The thread running through that list is edition. On the self-managed editions of every one of these forges, the native AI review story is thin or unstated, and the real decision is who runs the reviewer. Choose a hosted third party like CodeRabbit and a self-managed GitLab posts review comments under the coderabbit user, but GitLab itself must be 16.x. Choose PR-Agent and you control the model and the data path, but you also run the service.
The method that decides which list to trust
Every entry above names the edition it applies to and carries a date it was read. When a page does not state whether the feature works on the Data Center edition, this page says so and stops. An absence of documented feature is not proof of absence, and treating it as proof is how the confident GitHub-only lists get written.
The next time you search "best AI code review tools," add one word: your forge. Then check the vendor's own docs for the edition you run rather than a blog that assumed GitHub. If the vendor page is silent on the edition you run, that silence is the finding, and it is worth more than a list with no caveats.
Top comments (0)