Install both.They answer different questions.
Anthropic ships two security plugins and they are good. One reviews code as Claude writes it; the other runs a multi-agent scan of a repository and produces reviewed patches. Both read the source in your checkout. This page is about the questions that start after that sentence.
Nothing here argues that the official plugins are worse. The in-session plugin reduces what reaches your branch, and this tool answers the question you get asked afterwards — usually by somebody outside the engineering team.
Capability
The row that matters is the one you are about to need.
A tick in the SecAudit column is a claim this site can be held to. Where the honest answer is a figure rather than a tick, the figure is shown — and it is read from the same measurement the rest of the site publishes.
| Capability | Official plugins | SecAudit |
|---|---|---|
| Reviews code as Claude writes it | ✓ | — |
| Multi-agent repo scan → reviewed patches | ✓ | — |
| Audit a running site or API | — | ✓ |
| Authorization gate + scope.yaml for active testing | — | ✓ |
| Runs without Claude Code, without a paid plan, offline | — | ✓ |
| Published, reproducible detection score | — | 61.2% / 54.4% |
| SBOM, OpenVEX and EU CRA evidence pack | — | ✓ |
| Tools exposed to Codex, Cursor, OpenCode over MCP | — | 6 |
The two figures in the score row are recall on two external corpora — 61.2% F3 on Python and 54.4% recall on JavaScript, both scored by benchmarks this project does not own. Neither official plugin publishes a number, and their documentation says scans are nondeterministic, so the row is a difference in kind rather than in degree.
In their words
The limits below are quoted from Anthropic's own documentation.
Not paraphrased and not inferred. Each line is what the official docs say about their own scope, read on the date in the eyebrow above. If they change the product, this page is wrong until somebody updates it — which is the honest liability of writing about someone else's software at all.
Both quotations are from the Claude Code security documentation. Follow the link in the hero and read them in context; that is the point of quoting rather than summarising.
- “the review reads the source code in your checkout, not a running site or deployed service”This is the opening for a live-target track, and it is why this project has an authorization gate at all.
- “two scans of the same code can surface different findings”A deliberate property of an agentic scan, and the reason a deterministic tier can publish a reproducible number and an agentic one cannot.
What this does not do
Four things SecAudit deliberately is not.
A comparison page that only lists advantages is an advertisement. These are the four places where the answer is to use something else.
- An in-editor reviewerIt does not watch Claude write code. Install the official guidance plugin for that; the two do not overlap.
- A measured LLM tierThe optional model tier ships off by default and is not measured. A general-purpose model scores above this engine on the same corpus with no harness at all, and that is stated rather than hidden.
- A finder of business-logic flawsThe rules being broken are your product's and are not written down anywhere the analyser can read. Most of what this engine still misses is this, and the benchmark pages say so per class.
- A certificateThe compliance pack is input to a process, not proof of one, and the pack says so in its own disclaimer. Standards whose control text cannot be quoted are refused rather than mapped by number.
Each of these is stated somewhere else on this site too, with the measurement behind it. A limit that only appears on the page nobody reads is a limit that was not really disclosed.
Install the official plugins. Then install this one.
They review what you are writing. This audits what you have shipped, against a corpus somebody else labelled, and produces the document you hand over afterwards.