Questions

The questions that come up most, each with a short answer and where to read more.

Next to other tools

I already run npm audit, composer audit or Dependabot. What does this add?

Those tools read your lockfile and tell you which libraries have a published advisory. They do that well, and the skill runs them. They do not see a flaw in the runtime itself (PHP, Node.js), in a self-hosted tool (GitLab, Keycloak), a security release announced for next Tuesday, or whether a flaw is being exploited this week. stackalarm reads those, then the agent checks each one against your project.

Does it replace Trivy, Grype or a commercial scanner?

No. A scanner inspects an artifact and lists its known flaws. stackalarm reads the news and judges what concerns a project. The two meet on container images: write the SBOM with your scanner, and stackalarm adds what is exploited, what has a public exploit, and what is announced but not yet out.

Can I use the CLI without an AI agent?

Yes. The CLI works on its own and has no AI in it. You then pick the tags and do the sorting yourself. The news pages show the same news in a browser, with a feed for each tag.

Cost and accounts

Is it free?

Yes: MIT license, no account, no telemetry. Every source is public. The only thing that can cost money is the optional triage of the scheduled watch, which calls Claude with your own Anthropic API key, and only on the runs that found something new.

Do I need API keys?

No. Two optional keys help: a GitHub token lifts the cap of 60 anonymous requests per hour, and an NVD key makes the CVE searches about seven times faster.

When something is off

Sources fail with "rate limit reached"

GitHub caps anonymous requests at 60 per hour per IP address. Set GITHUB_TOKEN (any token, no permission needed): export GITHUB_TOKEN=$(gh auth token). The failed sources are always listed, so you know what is missing.

A run takes a minute or more

Without a key, the NVD allows one search every 6.5 seconds, and each tag and keyword is one search. Put fewer keywords, or set NVD_API_KEY. Responses are reused for an hour, so the second run is fast.

My technology has no tag

308 technologies have their own sources. Anything else works as a --keyword: the CERTs, the CVE databases and the press are searched for the word. To add its official feed, open a New source issue.

The agent does not use a command the docs describe

The skill is older than the CLI. They are installed separately: update both.

It keeps reporting a vulnerability that does not concern us

Record the decision in the profile, under ignore, with the reason and a review date. The item leaves the reports but stays listed with its reason, and comes back by itself if it starts being exploited.

Will the watch open an issue every 4 hours?

No. It remembers what it already reported and wakes up only for a new heads-up, alert or security release. With the triage, an issue is opened only when Claude finds the project concerned. Quiet runs open nothing.

Trust

What leaves my machine?

The tags, keywords and package names you give, sent as search terms to public databases. Never your code. The full table is on Privacy and security.

Can the agent be wrong?

Yes. It can misread a version or miss a condition of exploitation. That is why every finding names the file the version came from and links to the official advisory, why "Not checked" lists what it could not verify, and why it changes nothing without asking.

Does it change my project or my servers?

The CLI only reads. The skill offers fixes in the repository, one at a time, after you agree. It never pushes, deploys or connects to a server.