Claude Code

Security testing inside Claude Code, with the findings proven.

One slash command runs a structured application security assessment on your machine, on the Claude Code subscription you already have. Your repository stays local. Each finding says whether it was proven against your running application or only suspected.

npx myrqen@latest
Invocation
/myrqen auto
Account
not required
Model key
none

Claude Code already ships a security feature.

/security-review has been there since August 2025. It is free with any Claude Code access and needs no setup. If you have not run it, run it before you install anything.

It reads your code and reasons about it. That is useful, and it is a different job from this one.

What it does

/security-review

Static analysis of code and diffs.

Myrqen

A static pass, then each candidate exercised against the running application.

Can it prove a finding?

/security-review

No. It reasons about the code.

Myrqen

Yes, for the classes a safe probe exists for. Under two identities and an admin.

Does it say when it is unsure?

/security-review

Prose confidence.

Myrqen

A verification state the CLI owns: verified, strong_evidence, needs_review. An agent cannot promote its own finding.

Boundary on what it may touch

/security-review

Not needed. It touches nothing.

Myrqen

Exact-origin grants, a prohibited-action list, and quarantine for origins project text proposed.

Languages

/security-review

Whatever the model reads.

Myrqen

JavaScript and TypeScript; Python for FastAPI and Flask.

Artefacts

/security-review

Terminal output and PR comments.

Myrqen

HTML, JSON, Markdown and SARIF, written to your project.

Published accuracy

/security-review

None published.

Myrqen

Recall 0.250 on a held-out corpus, and 0.500 on an earlier one after the fixes it prompted, against a gate of 0.90. The static pass alone, scored one file at a time with no agent involved.

Works in Codex, OpenCode, Cursor

/security-review

No.

Myrqen

Yes. The same skill, thirteen agents.

Cost

/security-review

Free.

Myrqen

Free for local scanning.

If reading the code is enough, use the free thing. Myrqen is for when you want an authorization bug shown to you by fetching another account's record, and a written record of everything the assessment could not settle.

Install it in Claude Code.

One command. It detects Claude Code, shows you the plan, and writes nothing until you accept.

install

$ npx myrqen@latest

Detected coding agents
  ● Claude Code       Confirmed  2.0.14

Ready to link Myrqen
Claude Code · this project
  CREATE  .claude/skills/myrqen/SKILL.md
  MERGE   CLAUDE.md            existing content preserved

The command comes from the directory

The bundle installs as myrqen/, so it is invoked as /myrqen. Your CLAUDE.md is preserved: Myrqen owns only the region between its own markers, and myrqen unlink returns the file byte for byte.

Project or global

A project-scoped install writes into the repository, so the whole team gets the skill from the checkout. A global install writes into ~/.claude/skills/. myrqen link --scope project picks one.

Start your app. Then one command.

auto picks a depth from the size and shape of the project. low high xhigh ultra set it explicitly. Depth changes breadth, reasoning passes and validation depth. It never changes what is authorized.

in Claude Code

/myrqen auto

Claude Code works the phases. The CLI holds the contract.

  1. Discover

    Routes, identities, secrets and dependencies, read from your source.

  2. Probe

    Candidates exercised against your running application. Never production.

  3. Settle

    Each candidate verified, refuted, or left explicitly unsettled.

  4. Fix

    A remediation prompt and the test that should pass, handed straight back to Claude Code.

  5. Retest

    Run it again. The report says what closed and what came back.

The fix step is where running inside the agent pays off: Claude Code already has the file open, the context, and permission to edit. Copying a finding out of a dashboard into a chat window is the step this removes.

What a finding looks like.

Real output from the deliberately vulnerable fixture in the repository.

vuln-shop · MYR-004

MYR-004   high · verified
orders is read by identifier without checking who owns it
GET /api/orders/:id  ·  src/server.js:95

  Verified: signed in as user_a and requested /api/orders/102, a record
  listed for user_b. The response was HTTP 200 with a body of 93 bytes.

  Evidence · source · src/server.js:95
      if (!user) return json(response, 401, { error: "sign in first" });
      const id = Number(url.pathname.split("/").pop());
      const order = orders.find((candidate) => candidate.id === id);
      return json(response, 200, order);

Severity and verification are independent: "how bad if real" and "how sure it is real" are different questions. A candidate nobody exercised stays needs_review and says why. It is never promoted to make a report look better.

What it will not do.

  • Find everything. Recall is 0.250 on a corpus nobody had tuned against, and the same whether or not a finding is allowed anywhere in the range the case's author marked rather than within eight lines of the line they named. It is the static pass alone, scored one file at a time with no agent involved, against our own gate of 0.90. Every number, and the method.
  • Read Java, Ruby, Go, or a Python framework other than FastAPI and Flask. Java, Ruby and Go are not parsed; other Python frameworks are parsed, but their routes are not modelled.
  • Follow every path data can take. In JavaScript and TypeScript the static pass follows a value across modules only where a call resolves. Events, queues and dynamic dispatch are Claude's part of the work, and the engine does not claim them.
  • Anything destructive. At any depth. There is no flag.
  • Touch a host you did not name. Authorizing https://api.example.com authorizes nothing else.
  • Tell you that you are secure. It reports what it established and states what it could not reach.

Common questions.

Does Myrqen work with Claude Code?
Yes. It installs as a Claude Code Agent Skill and is invoked as /myrqen auto. myrqen link detects Claude Code, shows you the files it would write, .claude/skills/myrqen/ and a managed region in CLAUDE.md, asks, and then writes only those.
How is this different from Claude Code's built-in /security-review?
/security-review reads your code and reasons about it. Myrqen exercises each candidate against your running application under two identities and an admin, and records whether it was proven, supported by evidence, or left unsettled. /security-review is free and needs no setup, so run it first; use Myrqen when you want a finding demonstrated rather than described, and when you want an enforced boundary on what the agent may touch.
Do I need a separate API key or subscription?
No. Myrqen has no model of its own and never asks for a model provider credential. The reasoning comes from the Claude Code subscription you already pay for. There is nothing to buy to run a scan.
Does Myrqen upload my repository?
No. There is no code path that uploads your source tree. The only thing that can leave your machine is the finished report, and only if you answer yes to a question asked once per scan. Secret values never appear in a finding at all.
Which languages does it support?
JavaScript and TypeScript across Node, Bun, Deno and edge runtimes, and Python for FastAPI and Flask applications. Django and other Python frameworks are parsed, but their routes are not modelled, so data-flow checks do not start from them. Java, Ruby, Go, template languages and infrastructure definitions are not parsed at all. If your codebase is written mostly in anything else, use a mature SAST product instead.
Is Myrqen a SAST tool?
Partly. It has a static pass of 49 rules, but a static candidate is not reported as a finding until an agent settles it against the running application, and if nobody does, it stays labelled needs_review. It is also not software composition analysis: dependency checks cover a fixed watchlist of commonly used npm packages, from an offline advisory snapshot, and report a package only where your own code uses the affected export.
How accurate is it?
Precision is 1.000 and the false-positive rate on deliberately correct code is 0.000 on holdout-3 — but only after the fixes its blind run prompted, which burned it as a measurement. On holdout-7, a later corpus scored blind, the false-positive rate on deliberately correct code was 0.125 and 3 of four gates failed. Precision itself is not computable on it: precision is derived from exact class-name matches, and thirty independent authors named the defects in their own vocabulary. The safe-case false-positive rate is the precision signal that survives. Recall is 0.250 on a corpus nobody had tuned against, and the same whether or not a finding is allowed anywhere in the range the case's author marked rather than within eight lines of the line they named. It is the static pass alone, scored one file at a time with no agent involved, against our own gate of 0.90. A report with no findings is not evidence that an application is secure. The full method and every number are published.
Is Myrqen safe to run against my own app?
Destructive actions are refused at every effort level and there is no opt-in flag: no mass deletion, no request flooding, no credential spraying, no persistence. Anything outside your own machine needs an exact-origin grant, and Myrqen never tests production, it exercises the application running locally.
Is Myrqen free, and is it open source?
Local scanning, local reports in all four formats, and every finding at every severity are free and unlimited, with no account. It is not open source: the source is published under a source-available licence that grants reading and running, not redistribution.