AI Code Audit

Your codebase passed review. Will it survive handover?

Seven structural checks, run on your actual repository. Every finding comes with the file and line that produced it, and a fix order ranked by what breaks first in production. Read-only — your code is never executed.

Scan a repository free → See what each tier includes

Free scan: about 15 seconds, no signup, public repositories. The written report and the human-verified audit are paid.

Self-serve report

$297 one-time

You run it. You get the full write-up.

  • All seven checks scored, with file-level evidence
  • Full written report, not just the score card
  • Fix order ranked by what breaks first in production
  • Written review of your two weakest checks
Get the report — $297 →

Human-verified audit

$2,500 7 days

We run it, read it, and defend the findings.

  • Everything in the report, plus false-positive removal
  • A person ranks the findings and writes the fix order
  • Recorded walkthrough of your code, on video
  • Free re-scan after your fixes land
Book the audit — $2,500 →

Not sure yet? The free scan already gives you the score and the evidence lines. The paid report adds the written analysis, the ranking and the fix order.


What the seven checks measure

These are not style rules. Each one asks a question a buyer, a CTO or a client will ask at handover.

CheckWeightThe question it answers
C1 · Scope containment18%Does the code stay inside its own boundaries, or does one module reach into everything else?
C2 · False-green detection18%Can your test suite actually fail? Tests that pass no matter what are worse than no tests.
C3 · Drift containment14%Do configuration, prompts and version numbers disagree between places they should match?
C4 · Memory integrity10%Is accumulated state bounded, or does it grow until something falls over?
C5 · Delivery gate16%Is there a CI or pre-commit gate that blocks a bad merge, and is the dependency tree locked?
C6 · Complexity control12%Are shipped files of a size a new maintainer can read, or are they 30,000-line engines?
C7 · Regression surface12%How much of the codebase can be broken by a change somewhere else?

Score bands: 85+ low risk · 70–84 medium · 50–69 high · below 50 severe.

Pricing, and why it is $297

Comparable fixed-fee code audits publish list prices between $299 and $499 for a written report on one codebase. Automated-only readings start around $49; audits written for an investor or a buyer run $1,500–$7,500.

We sit at the bottom of the middle band on purpose. The $297 tier is self-serve: you run the scanner, so there is no human hour inside the price. The human-verified tier is $2,500 — seven days, a person on the findings, a recorded walkthrough, and a re-scan after your fixes. Both prices are fixed. There is no hourly billing and no quote that grows after you commit.

What this is not. It is not a penetration test, not a vulnerability scan, and not a security certification. It does not run your code, does not test behaviour, and does not find every bug. It measures whether the repository has the structural evidence a delivery is usually accepted on — and it will say so plainly when a check cannot be assessed at all.

Public audits you can read before paying

Three mature open-source AI repositories, scored with the same seven checks. Same method, same evidence format, no editing.

pydantic-aisee report → openai-agents-pythonsee report → autogensee report →

Questions

What exactly is an AI code audit?
A read-only review of your repository against seven structural checks that decide whether a codebase survives handover. You get a score, the evidence line behind every finding, and a fix order.
What do the seven checks actually look at?
File layout and directory scope, test files that cannot fail, configuration and prompt drift between versions, memory and state handling, CI and pre-commit gates plus lockfiles, oversized shipped files, and the surface where one change breaks unrelated behaviour. Structural, not stylistic.
Is it automated or human-reviewed?
The free scan and the $297 report are automated — the same seven checks, run on your repository, evidence lines printed. The $2,500 tier is human-verified: a person reads the findings, removes false positives, ranks them, and records a walkthrough.
How long does it take?
The free scan returns in about fifteen seconds and the written report is immediate. The human-verified audit is seven days from repository access, with a free re-scan after your fixes.
Do you need access to my code?
For the free scan and the $297 report, only a public repository address. For the 7-day audit, read-only access to a private repository or an archive. Your code is never executed, and nothing is retained after the deliverable.
How is this different from a penetration test?
A penetration test attacks a running application from outside. This reads the source from inside and measures structural risk. It is not a vulnerability scan and does not claim to be one.
How is it different from SonarQube or a linter?
Linters flag patterns inside files. These checks look at the repository as a delivery system: whether the test suite can fail, whether CI blocks a bad merge, whether configuration drifts between environments. A clean linter score and a bad audit score is a normal combination.
What if my codebase is huge?
The scanner accepts archives up to 120 MB and skips dependency and build directories. For very large monorepos the human-verified audit scopes to the services you are about to hand over.
My app is already in production. Is this safe?
Yes. The scan is read-only and static: it downloads a copy of the source, reads it, and never runs it. No downtime, no deployment, no write access.
Do you sign NDAs and work on private repos?
Yes. An NDA is available on request for the human-verified audit, and private repositories are supported through read-only access.
Do I own the report?
Yes, in full. Forward it, attach it to a client handover, or give it to another developer to act on. Nothing is gated behind an account.
Can you fix what you find?
No. We produce the findings, the evidence and the fix order; we do not take on remediation. Deliberate: an audit that ends in a repair quote is a sales document, not an assessment.
Which stacks and AI builders do you cover?
The checks are structural rather than syntactic, so Python, TypeScript, Go, Rust, PHP and Java score the same way — whether the code came from Cursor, Claude Code, Copilot, Bolt, Lovable, v0, or a human team.
What if the score comes back bad?
A score is a starting point, not a verdict. A low score means the repository has no structural evidence that it can survive handover. The report tells you which of the seven is costing you most, and in what order to fix.