CodeRabbit Documentation - AI code reviews on pull requests, IDE, and CLI
Skip to main content
CodeRabbit scans your repository for well-known AI coding assistant configuration files and uses their content as review criteria. If your team already writes instructions for tools like Cursor, Claude, or Windsurf, CodeRabbit picks those up automatically and enforces the same standards during code review.

​
Supported files

The following file patterns are detected by default:
File names are case-sensitive. A file named claude.md is not matched by the **/CLAUDE.md pattern.

​
How scoping works

By default, a guideline file applies to the directory it lives in and all of its subdirectories. CodeRabbit does not apply guidelines from one part of your repository tree to unrelated paths unless you explicitly map the guideline to source-file patterns with applyTo in a filePatterns object entry. Examples:
  • CLAUDE.md at the repository root → applies to all files
  • src/frontend/CLAUDE.md → applies only to files under src/frontend/
  • src/backend/.cursorrules → applies only to files under src/backend/
This directory-scoped behavior means you can maintain separate, purpose-fit guidelines for different areas of a monorepo without them interfering with each other. When directory placement does not match the files a guideline should govern, use the object form of filePatterns to define the mapping explicitly.
Without an explicit applyTo mapping, guidelines in a documentation or tooling directory will not affect code review unless the reviewed files live inside that same directory tree.
A common mistake is adding guideline file names (for example CLAUDE.md) to path_instructions. This tells CodeRabbit to review those files as changed code, not to use them as guidelines. Use filePatterns instead (see below), or rely on auto-detection.

​
Managing applied guidelines

The CodeRabbit UI shows the applied coding guidelines for a repository. Repository admins with write access can view applied guidelines and disable them. Select one or more applied guideline rows and click Trash selected to stop those rendered guidelines from being applied in future reviews. Trashing a guideline does not delete the source guideline file from your repository. It suppresses the rendered path pattern and instruction text that CodeRabbit applied for that repository. If the source guideline changes enough to render as a different instruction, it can appear again as a new applied guideline. To turn off code guidelines entirely for a repository, set knowledge_base.code_guidelines.enabled to false.

​
Adding custom file patterns

If your team stores coding standards in files that are not in the default list, you can extend the detected patterns by setting knowledge_base.code_guidelines.filePatterns in your .coderabbit.yaml. Each entry can be either a glob string or an object with files and applyTo fields. A string entry locates guideline files and scopes each matched file to its containing directory and subdirectories. An object entry uses files to locate the guideline document or documents and applyTo to define the source-file glob or globs they govern. This explicit mapping lets you store guidelines outside a source tree and apply them to any matching files. Custom patterns are added on top of the defaults—they do not replace them.
.coderabbit.yaml
Glob patterns follow the same syntax used elsewhere in CodeRabbit configuration. The ** wildcard matches any number of path segments. In the configuration editor, the Files and Apply To fields correspond to the files and applyTo properties in an object entry.

​
Guidelines from another repository

| A string entry can name another repository as the source of your guideline files. Write it as repo:path to reuse the reviewed repository’s organization, or as owner/repo:path to name the organization explicitly. On GitHub, the source and reviewed repositories must be in the same organization; on GitLab, they must be in the same top-level group. The path can be an exact file path or a glob, and it can be mixed freely with unqualified entries that locate files in the repository being reviewed. Entries that name no repository keep selecting files from the reviewed repository, exactly as before, following the directory-based scoping described above.
.coderabbit.yaml
At review time, CodeRabbit reads matching files from the source repository’s default branch. Cross-repository guideline content is not pinned to a per-file revision, so later reviews can use updated content from that branch. Object entries support cross-repository sources too: a files value may use the same repo:path or owner/repo:path form, and the entry’s applyTo scopes the resulting guidance to matching paths in the repository being reviewed. This lets a centralized standards repository declare which files each guideline document governs — for example, a Python style guide that applies only to **/*.py in every consuming repository:
.coderabbit.yaml
Guidelines selected from another repository through string entries apply globally to the reviewed repository as review instructions; an object entry’s declared applyTo narrows that guidance to the paths it names. CodeRabbit does not treat cross-repository guidelines as repository-specific learnings. Both entry forms apply when set through committed configuration files, the web interface, or central configuration. Pull request description overrides do not yet honor object-form entries for guideline selection — use string entries there.

​
Eligibility and access

  • Installation read access is required. The CodeRabbit installation must be able to read the source repository so it can resolve and load matching files.
  • Administrator-approved sources skip per-user checks. Sources set through trusted organization, central configuration, or target-branch configuration apply to every review, including for authors who cannot read the source repository directly.
  • Not available on self-hosted CodeRabbit. This feature requires CodeRabbit Cloud. That requirement is independent of whether your GitHub or GitLab instance itself is self-managed.

​
GitHub

  • Same organization only. The source and reviewed repositories must belong to the same organization. CodeRabbit rejects references to another organization, and ignores an entry that names the reviewed repository itself because local discovery already covers it.
  • Author and pusher access is checked. When a source is introduced through configuration on a collaborator’s pull request branch, or through a pull request description override, CodeRabbit confirms the pull request author can read the source repository and, if someone other than the author pushed to the branch, that they can read it too.

​
GitLab

  • Same top-level group only. The source and reviewed repositories must belong to the same top-level group. CodeRabbit rejects references outside that group and ignores an entry that names the reviewed repository itself because local discovery already covers it.
  • Reporter-or-higher membership is required. The merge request author and any distinct human account that triggers the review must have effective Reporter, Developer, Maintainer, or Owner membership in the source project. Guest membership does not qualify. Effective membership includes access inherited from a group or subgroup.
  • Usernames must match exactly. CodeRabbit resolves the exact GitLab username for an access check; prefixes and near matches do not qualify.
  • Separate actor checks apply only when needed. CodeRabbit skips the separate triggering-account check when the merge request author triggers the review or when CodeRabbit starts an automatic or command-triggered review.
  • Project visibility does not replace membership. Public or internal visibility without effective membership does not grant access to the source project for this feature.
Access checks fail closed. If CodeRabbit cannot confirm an identity or the required access on either provider, it omits that source and continues the review with the remaining guideline context instead of failing the review.

​
Limits and validation

  • Entries. You can configure at most 50 entries that name another repository.
  • Expanded files. CodeRabbit reads at most 50 files per review after those entries’ globs expand, counted across all source repositories.
  • Entry length. Each entry that names a repository can be at most 512 characters.
CodeRabbit rejects absolute paths, parent-directory traversal, backslashes, malformed repository names, and entries with extra colons. Only text and documentation files are read. A file is eligible when it has no extension, is a dotfile such as .cursorrules, or uses one of these extensions: .md, .mdc, .txt, .rst, .json, .yaml, .yml, .csv, .log, .conf, .cfg, .ini, .toml, .xml, .html, .tsv, .ndjson, .jsonl, .adoc, .tex, .asc. Source files such as .ts are skipped even though they are plain text.

​
Data retention

Local guideline discovery is stateless. Cross-repository guideline sourcing stores content-derived guideline digests and is excluded for organizations that set knowledge_base.opt_out. See Opt out of data retention.

​
Adding guidelines to a single pull request

Guideline sources can also be selected per pull request, which is useful when the standards that matter depend on what a pull request changes. Add a configuration override to the pull request description: the @coderabbitai configuration override line as plain text, immediately followed by a fenced YAML block.
Guidelines added this way apply only to that pull request, and only in addition to the guidelines your configuration already defines. They cannot replace or disable the configured set, so your organization’s standards always still apply. Because the selection lives in the pull request description, it can be written by hand or set programmatically. Tooling that opens or updates pull requests can choose the guideline documents that match the area of the codebase being changed, keeping the committed configuration to the standards that apply everywhere. This requires a pull request from a branch in the repository itself, opened by a repository collaborator, member, or owner. Guideline entries are not accepted from a fork’s description: CodeRabbit rejects the block that contains them and notes it in the pull request. Access to entries added this way is checked the same way as other collaborator-introduced sources — see Eligibility and access above. For the full set of settings a description override accepts, see configuration override.

​
Guideline status in the review summary

The review summary reports the outcome of collecting cross-repository and pull request description guidelines:
  • ✅ Applied guideline sources (N). This confirmation appears when CodeRabbit collects one or more guideline files. The listed entries are the actual collected files in owner/repo:path form. Repeated entries are deduplicated so each collected file appears once.
  • 🔒 Skipped guideline sources. This note identifies sources that CodeRabbit could not collect, including unreachable or invalid sources. CodeRabbit continues the review with local guidelines and other reachable sources, so a skipped source does not block the review.

​
Configuration reference

.coderabbit.yaml
Setting filePatterns to an empty list ([]) keeps the default patterns active. To disable code guidelines entirely, set enabled: false.

​
What’s next

Knowledge base overview

Learn about all knowledge base capabilities: learnings, code guidelines, and linked repositories.

Learnings

Teach CodeRabbit your team’s review preferences through natural conversation.

Configuration reference

Full reference for all knowledge_base configuration options.