Documentation

Cerberus Cloud

Automated AI code review on every pull request — no YAML, no API keys, no configuration. Install the GitHub App and the swarm gets to work.

Quick start

  1. 1

    Install the GitHub App

    Go to the GitHub Marketplace listing and click Install it for free. Choose to install on all repositories or select specific ones.

  2. 2

    Choose a plan

    Select the plan that matches your team's PR volume. All plans include a 14-day free trial. Billing is managed through GitHub Marketplace — no separate account needed.

  3. 3

    Open a pull request

    Cerberus automatically triggers on every new PR and every push to an open PR. Review feedback appears as a GitHub check run and inline PR comments within a few minutes.

That's it. No YAML workflow files to write. No API keys to manage. No configuration required. Cerberus handles everything.

How it works

When a PR opens or receives a new commit, Cerberus:

  1. Fetches the PR diff, file list, and commit metadata from GitHub
  2. Builds a structured context bundle — no full repository clone, no source code stored
  3. Launches six specialist reviewers in parallel, organized into three escalating waves
  4. Each reviewer emits a structured verdict: PASS, WARN, or FAIL with specific findings
  5. The council verdict is synthesized and posted as a GitHub check run and PR review comment
  6. If the verdict is FAIL, the check run fails — blocking merge until resolved or dismissed

The council

Six specialist reviewers, each with a singular mandate. Breadth comes from composition, not from asking one model to think about everything at once.

APOLLO·Correctness

Logic errors, edge cases, race conditions, and semantic bugs.

SENTINEL·Security

OWASP vulnerabilities, secret leaks, auth bypasses, and injection risks.

VULCAN·Performance

Big-O complexity, N+1 queries, unnecessary I/O, and latency hotspots.

ATHENA·Architecture

Coupling, modularity, pattern consistency, and structural integrity.

ARTEMIS·Maintainability

Readability, naming, documentation, and test coverage gaps.

CASSANDRA·Testing

Test completeness, edge case coverage, and assertion quality.

Wave escalation

Reviewers run in three waves. Each wave only proceeds if the previous wave exits clean — keeping compute proportional to actual code risk.

Wave 1 — FlashFast models

Rapid static analysis. Catches obvious security issues, syntax problems, and known vulnerability patterns. Completes in seconds.

Wave 2 — StandardStandard models

Cross-file semantic analysis. Logic flow, algorithmic complexity, and architectural alignment.

Wave 3 — ProReasoning models

Deep reasoning. High-level architecture review, system-level race conditions, and long-horizon correctness.

Verdicts

Each reviewer emits one of three verdicts. Only findings above the confidence threshold count toward the final decision.

PASS

No significant issues found. The check run succeeds.

WARN

Issues found but not blocking. Posted as PR comments. Merge is allowed.

FAIL

Blocking issues found. The check run fails. The PR cannot merge until issues are resolved or the review is dismissed by a repo admin.

Each finding includes: severity, category, file path, line number, title, description, and a suggested fix.

Plan limits

A review run is one council session — one PR, one swarm. Not per-seat, not per-line.

PlanReviews / monthPriceSeats
Free10$0Unlimited
Starter100$29/moUnlimited
Pro300$59/moUnlimited
Team800$119/moUnlimited
Enterprise10,000Contact usUnlimited

When the monthly quota is reached, reviews pause until the next billing cycle. Upgrade anytime through GitHub Marketplace.

FAQ

Does Cerberus store my source code?

No. PR diffs are processed in memory during the review and discarded after the verdict is generated. We do not retain source code. See our Privacy Policy for full details.

Can I use Cerberus Cloud on private repositories?

Yes. Cerberus Cloud works on public and private repositories. The GitHub App requests only the permissions required: read access to pull requests and code, write access to check runs and PR comments.

Which languages does Cerberus support?

Any language. The reviewers analyze code structurally and semantically — they are not linters and do not depend on language-specific parsers.

How long does a review take?

Wave 1 typically completes in under 30 seconds. A full three-wave review on a mid-size PR takes 1–3 minutes. Large diffs may take longer.

Can I configure which reviewers run?

Not in Cerberus Cloud — all six reviewers run on every PR. Full configuration is available in the open-source self-hosted version.

What happens when I hit my monthly quota?

Reviews pause gracefully. Open PRs are not blocked — the check run is skipped with a quota message. Reviews resume at the start of your next billing cycle, or immediately on upgrade.

Is there a self-hosted open-source version?

Yes. Cerberus on GitHub is a GitHub Action you run in your own pipeline. Bring your own OpenRouter API key. Full control over models, prompts, and thresholds.