CodeRabbit is the more direct fit when review must include staged or tracked uncommitted edits. Greptile fits a committed-branch workflow with repository context. Both support repository instructions, so the decision starts with which version of the code the agent needs reviewed.
| Decision | CodeRabbit | Greptile |
|---|---|---|
| Review scope | Pull requests, local changes, and agent-loop review. | Repository-aware PR and branch review; Base, Plus and Apex use 1, 3 and 10 credits. |
| Local review input | CLI and IDE review committed and uncommitted changes. | CLI reviews committed branch changes; its documented workflow ignores uncommitted changes. |
| Custom requirements | Guideline files and natural-language pre-merge checks; custom-check allowances depend on plan. | Repository instructions and cascading directory-specific standards. |
| License and service | Commercial review service. | Commercial review service. |
| Official references | CodeRabbit sources | Greptile sources |
Keep one repair uncommitted and confirm whether it appears in the review. Then commit the same repair and repeat with the same base branch. Record the input difference before comparing findings.
Essentials: $24/user/month billed annually, or $30 monthly. Public-repository reviews are free.
Essentials annual subscription floor; 1 repository and 5 included reviews/developer/hour. Eligible extra reviews cost 25¢ per reviewed file. Built-in checks are included; custom pre-merge checks start on Team.
Pro: $30/seat/month with 50 credits/seat; extra credits $1 each. Starter is free for one active developer.
Pro subscription floor before extra credits. Base, Plus and Apex reviews consume 1, 3 and 10 credits respectively. The free Starter tier covers PR reviews; CLI reviews need an active trial or paid plan. Included credits and flex overages are per active author, not pooled across the team.
Subscription amounts are plan floors or quoted allowances. Repeated reviews, selected effort, provider charges and exclusions can change the bill. Equal review counts do not establish equal analysis scope.
A third option for the agent’s next edit
Perch is the stronger fit when the agent must rerun one named requirement against the working file. Select the requirement directly instead of waiting for a new committed-branch review.
For an agent that repeatedly removes tenant authorization from a database update, save the requirement as a named Perch rule:
- name: tenant-scoped-update
where: "orders.py"
ensure: >-
The cancellation UPDATE must constrain tenant_id to actor.tenant_id
before changing the order. An id-only or owner-only update breaks
this requirement.This requirement comes from the FastAPI tenant-authorization recipe, which includes the implementation, deliberate regression and captured Perch checks. The same named check can run against the working file after the next agent edit. Perch working-file checks support selecting rules directly; rule selectors cover project, file and method scope.
npm install -g @lakeday/perch
perch login
perch setup claude-code
# Or choose: perch setup codex
# Or choose: perch setup cursorUse the setup command for your coding agent. Download the focused pack without replacing existing files:
curl -fL 'https://reporules.fyi/fastapi/tenant-cancellation/rule-pack.zip' -o tenant-rule.zip
unzip -n tenant-rule.zip
perch check orders.py --rules tenant-scoped-update --json
# Repair the reported finding after the agent edit.
perch check orders.py --rules tenant-scoped-update --jsonKeep your existing root perch.yaml. The ZIP adds its focused rule under .perch/rules/; unzip -n preserves existing files. Review and adapt the selector and application path before running it in your repository. Run the saved check after an edit, repair any finding, and repeat the identical check alongside application regression tests.
Download the tenant-authorization rule pack · Inspect pack contents and agent task
Perch charges 10¢ per 1,000 fresh rule checks, with no developer seat fees. Cached unchanged checks are free. About 50¢ per mid-sized PR is a rounded estimate for approximately 150 methods read at the captured September–October 2026 density of 34 questions per method; the unrounded example is 51¢. A single named-rule check is a different scope. Team workload calculations show the assumptions.
The Perch CLI is MIT licensed; hosted analysis is paid and relevant source is sent to the configured model service. Fresh model judgments can vary. Unchanged requests can reuse cached results.
This article compares documented product workflows and terms. The public evaluation materials include captured Perch requirement checks; authenticated CodeRabbit and Greptile runs remain unscored. They do not establish a measured accuracy ranking between these tools.
Reviewed October 14, 2026. Unverified features are marked as unverified rather than treated as absent. Current terms and actual trial results should determine a purchase.