audit/bench
AI code review, before it ships

AI coding assistants ship fast.
Make sure what they ship is safe.

audit/bench combines LLM reasoning with static analysis to catch the security holes, logic bugs, and framework misuse that traditional linters miss — on a single file, a pull request, or a whole repository.

audit/bench
await supabase
.from("invoices")
.update({ status: "sent" })
.eq("id", invoiceId);
Critical · Security
No auth check — any caller can mark any invoice as sent.
01 — Catches it
5
Review lenses per audit
3
LLM providers to choose from
11+
Frameworks understood
2
Git providers, native PR/MR review
Works with the stack your team already runs
ReactNext.jsNode.jsNestJSPython/FastAPIDjangoLaravelSpring BootSupabaseDenoFirebase
The problem

Modern AI coding assistants generate code quickly but frequently introduce hidden bugs, security issues, and framework misuse. Traditional linters catch syntax problems — not business logic or intent.

// one audit engine, wired into everywhere you ship.

Reviews. Gates. Fixes. Reports. Same findings whether it's a file, a PR, or a whole repository.

audit/bench
$ auditbench scan .
Scanning 42 files across src/, api/, lib/…
12 files flagged for AI review — 30 skipped, no risk found
api/webhooks/stripe.ts — signature not verified before processing
Verdictdo_not_ship
01One engine: files, PRs, whole repos
For engineering organizations

Built for how teams actually ship

Beyond one-off audits: review the code your team merges, track quality trends over time, and wire it into the pipeline you already have.

Pull & merge request review

Review a GitHub PR or GitLab MR scoped to just the changed lines — findings post back as inline review comments, a summary, and a merge-blocking status check.

Team analytics dashboard

Security, performance, and technical-debt scores trended over time, plus a breakdown of your most common findings and riskiest files.

CLI for CI pipelines

Run `auditbench scan` or `auditbench audit` from a pipeline step, a pre-commit hook, or a terminal — same engine, same findings, scriptable output.

Role-ready plans & quotas

Daily and monthly AI-audit limits per plan, usage-based — not a flat seat count — so cost scales with what a team actually reviews.

Where you already work

Lives inside the review you're already doing

Not another tab to check. Findings, gates, and answers show up directly on the PR or MR.

Inline PR & MR comments

Findings land as real review comments on the exact changed lines — GitHub review threads or GitLab discussions — not a dashboard you have to remember to check.

Auto-generated summary

Every review posts a plain-English walkthrough alongside the inline comments — verdict, finding counts, and a link to the full report.

Merge-blocking quality gates

A commit status check reports pass/fail on every PR and MR. Wire it into branch protection and stop shippable-looking regressions before they merge.

README score badge

A live, always-current badge for your README showing the verdict of your most recent scan.

Dependency vulnerability scanning

npm audit for Node, OSV.dev for Python — known-vulnerable packages surface automatically, no extra tooling to install.

Conversational PR chat

@-mention the bot in any PR or MR thread and it replies in context, using the diff to answer follow-up questions.

Cost control

Engineered to control AI spend

AI credits, not raw request counts. A three-stage pipeline keeps the LLM off the critical path until the code actually warrants it.

01
Free local checks

ESLint, TypeScript diagnostics, complexity, formatting, and secret scanning run first — no AI, no cost.

02
AI on flagged code only

Only the functions Stage 1 flags as risky go to an LLM, with just the relevant code and types — not the whole file.

03
Cached by content hash

Re-scanning unchanged code — the common case across repeated repo scans — costs nothing and returns instantly.

Plans

Start free, upgrade when you need more

Compare all plans →Need a custom rollout? Talk to sales.

Review your first repository free

No credit card required. Connect GitHub or GitLab, or upload a .zip, and see what an audit finds in your own code.