Skip to main content
Two commands, then your agent does the work. Nothing is uploaded, no account is required, and the whole thing works with the network off.
Myrqen is only for projects and targets you own or are authorized to test. Run it against something of yours. Pointing it at somebody else’s application is not what it is for, and the authorization gate refuses any origin you have not granted by name.

1. Install

That is myrqen link: it detects the coding agents installed on this machine, prints exactly which files it would write, asks, and then writes only those. Existing instructions are preserved, a second run changes nothing, and myrqen unlink removes only Myrqen’s own content. You need Node.js 20.12 or newer and one of thirteen supported coding agents — or no agent at all, if you are content to drive the CLI yourself. Confirm what this machine can assess:
doctor reports capability rather than alarm. “No account is connected” is info, not a problem: scanning is local and complete without one.

2. Start the application you want assessed

The scan is far more useful against something that is running, because that is what turns a static candidate into a verified finding.
Static-only findings are still reported. They arrive as needs_review and the report says so rather than implying they were proven.

3. Run the assessment

auto inspects the project’s size and surface and resolves a concrete effort profile, then prints the session context: the local report id, the resolved effort and why, the local targets it found, the external origins it will not touch without an explicit grant, and the artifact paths it will write. --sync no keeps everything local. Without it, and on a linked device, Myrqen asks once whether to sync this scan — see Cloud sync. Read the context back at any time:

4. Read the report

This validates every finding, then writes four files:
In the HTML report, filter by severity and by verification — they are independent — expand a finding for evidence, root cause, remediation, and the verification step, and read Coverage and limitations at the bottom, which names what was not assessed. The HTML is genuinely self-contained, so it opens with the network off:
Add .myrqen/reports/, .myrqen/sessions/, .myrqen/current-session, and .myrqen/local.json to the project’s .gitignore.
A report with no findings is not a clean bill of health. Its recall is 0.250 against a gate of 0.90 on a corpus nobody had tuned against — measured on 32 cases, so the 95% interval is 0.133 to 0.421. Treat the interval as the figure: on a sample this size a difference below about 0.346 is not measurable, and the four blind runs recorded so far are not distinguishable from one another.That run also found what the limit is. Where the dangerous value reaches a language or runtime primitive — fetch, a filesystem path, a shell, eval — recall was 0.875. Where it reaches a framework’s own database or identity call, it was 0.042, on five stacks the scanner had no specific support for. It is much better at a defect whose sink is JavaScript than at one whose sink is your framework’s vocabulary.On the deliberately-correct half of that corpus, 4 of 32 files drew a finding. Two were rules asserting a defect that is not there, which is a mistake; two were rules marking something needs review, which is a question. Myrqen ships those on purpose — the alternative is silence on the shapes a static reading cannot settle, and silence is how an unlimited sign-in endpoint came to produce no finding at all. The two are reported separately on Benchmarks rather than added together, so a question is not counted as a false claim and a false claim is not hidden behind one. Both figures are the static pass alone, scored one file at a time with no agent involved; what the host agent adds on top has never been measured against a corpus. It does not parse Java, Ruby or Go at all, models Python routes only for FastAPI and Flask (Scan a Python app), and follows JavaScript data across a module boundary only where a call resolves. All of that is stated in every report’s coverage notes and measured in Benchmark.

Reproduce a deterministic run

The steps above involve a language model, so no two runs are identical. The repository ships a fixture with machine-readable ground truth and a scripted stand-in agent, which together produce the same result every time — this is what the benchmark gate and every published number are computed from. It needs a source checkout, because the fixture and the harness are not part of the npm package:
The simulator drives exactly the CLI contract the Agent Skill uses, and reports how many deliberately unsafe plans the safety policy refused, which external origins were denied authorization, and how many findings were recorded, merged as corroboration, or rejected.
fixtures/vuln-shop is a Myrqen-owned application with deliberate defects, described in fixtures/vuln-shop/ground-truth.json. It binds to loopback only and keeps everything in memory. Every credential in it is a fake canary used to prove that redaction works, and none of them reaches any artifact:
Never deploy it anywhere.
Two of the fixture’s external origins are named by an injected instruction block in its README.md that tells the scanner to widen scope. Those origins are quarantined: the authorization gate refuses them even with --grant, because a human approving a target that injected text authored is not meaningful consent.

Next steps

Run it with your own agent

The exact contract the agent works inside, phase by phase.

Understand the safety model

Read this before you point Myrqen at anything that is not on your own machine.

Fix a finding and retest

The remediation prompt, the applied change, and the honest verification record.

Connect an account

Optional. Adds live progress, retention, and one share URL that renders differently per recipient.