ghostsecurity/skills: AppSec Skills for Claude Code

What problem ghostsecurity/skills solves, and for whom

Application security tooling has a handoff problem. A scanner runs in CI, produces a report, and someone triages it days later with no memory of the code. Ghost Security's repository attacks that handoff by moving the checks into the place where the code is already being read: the Claude Code session.

The repository is described as a plugin marketplace for Ghost Security's AI-native application security skills. That word, marketplace, is the important one. The repository does not ship a scanner binary. It ships skill definitions that Claude Code loads, and those skills drive security analysis from inside the agent.

The audience is narrow and specific. It is for engineers and security teams who already use Claude Code as their working environment and want repository context, dependency analysis, secret detection, code scanning, reporting, and dynamic validation available as commands rather than as a separate product to log into. If your team does not use Claude Code, most of this repository is inert to you.

The eight skills and how they chain together

The README lists eight skills. They are not independent tools; they read as a pipeline with a shared context layer underneath.

ghost-repo-context builds what the README calls shared repository context: business criticality, sensitive data, and a component map. Everything downstream depends on this. A dependency finding is only interesting if the agent knows whether the affected component handles sensitive data, and that judgement comes from here.

Three scanners sit on top of that context. ghost-scan-deps performs exploitability analysis of dependency vulnerabilities, which the README groups under SCA. ghost-scan-secrets does context assessment of detected secrets and credentials, so a hardcoded token in a test fixture and one in a production config are not treated the same. ghost-scan-code is described as AI-powered detection of code security issues, the SAST layer.

Two skills sit above the scanners. ghost-report combines results across all scan results into one report. ghost-validate performs dynamic validation of findings against a live application, which the README labels DAST. ghost-proxy is not an analysis skill at all: it is an HTTP proxy that exists to serve ghost-validate, which tells you the dynamic path needs a network intermediary.

The eighth skill, ghost-exo, is different in kind. It builds, improves, and debugs workflows on the exo agent orchestration platform. That is infrastructure for the agent itself, not a security check, and it is the clearest signal that this repository is as much about running agents as about scanning code.

Installing the plugin and running a first scan

Installation happens through Claude Code's own plugin commands. The README gives a command-line form and an in-session form; both add the same marketplace and install the same plugin.

From a shell, the three commands below add the marketplace, install the Ghost plugin from the ghost-security marketplace, and start Claude Code. The README notes that you currently need to restart Claude Code for the plugin to load, so expect the plugin to be absent until you do.

claude plugin marketplace add ghostsecurity/skills
claude plugin install ghost@ghost-security
claude

The equivalent inside a running Claude Code session uses slash commands instead:

/plugin marketplace add ghostsecurity/skills
/plugin install ghost@ghost-security

After the restart, the skills become available to the agent. A sensible first use is ghost-repo-context, because the other skills consume what it produces. The README does not print the invocation syntax for individual skills, so the exact command form is something to confirm in the plugin's own README at plugins/ghost/README.md rather than something to guess. What you should see, based on the skill description, is the agent building a picture of the repository's criticality, sensitive data, and component map before any scanning starts.

Where this design breaks down

The dependency on Claude Code is the first limitation, and it is not incidental. Every skill here is a Claude Code plugin. There is no documented CLI entry point, no CI action, and no server mode. If you want scanning to run on every pull request without a human in an agent session, this repository does not describe how.

The second limitation is release discipline. The repository has no retrieved releases, so there is no version to pin, no changelog to read before upgrading, and no tag to roll back to. The README itself says you currently need to restart Claude Code for the plugin to load, which is a rough edge in the install path rather than a documented, versioned process.

The third is the dynamic validation path. ghost-validate runs against a live application and depends on ghost-proxy. That means a running target, network reachability from wherever the agent executes, and a proxy in the middle. Teams that only have static source access cannot use that skill at all, and the README does not describe what ghost-validate does when the target is unreachable.

There is also a scope question worth naming. The README describes what each skill covers but does not document detection rules, supported languages, or how findings are scored. For a security tool, that is a real gap: you cannot reason about false positive rates from the repository alone.

How it compares to wiring scanners into CI

The obvious alternative is the conventional stack: Semgrep or CodeQL for static analysis, a secret scanner, and a dependency scanner, all invoked from a CI pipeline with results posted to the pull request. That approach has different properties, not better ones across the board.

The CI approach is deterministic and reviewable. You pin a scanner version, the ruleset is a file in the repository, and the same commit produces the same findings. It runs without an agent, without a model, and without a session. Its weakness is context: a CI scanner does not know that a particular service is business critical or that a config file holds production credentials, so it reports everything at the same severity.

ghostsecurity/skills inverts that. It trades determinism and pipeline integration for context, using ghost-repo-context to make downstream findings relative to the repository's actual importance. That trade is the whole point of the project. It is also why the two approaches are not substitutes: a team could reasonably keep CI scanners as the gate and use these skills for triage and validation, where the extra context changes the decision.

Maintenance cost, licensing, and what to check before adopting

The repository's last push was on 2026-09-03, twelve days before this writing, and it is not archived. On the evidence available, the project is being worked on now. That says nothing about the cadence going forward, and with no releases retrieved there is no version history to infer a rhythm from.

Upgrade cost is therefore the thing to plan around. Because installation is a marketplace add plus a plugin install, and both pull from the main branch, you get whatever is current. There is no documented pinning mechanism in the README. If your organisation requires approved versions of security tooling, that is a mismatch you need to resolve before deployment, not after.

The licence is Apache-2.0, and the repository carries a LICENSE file at its root. Apache-2.0 is a permissive licence with an explicit patent grant and requires preservation of notices. What it does not resolve is the data question: these skills send repository context to a model through Claude Code, and the repository does not document what leaves the machine. That is a question for your own review, not one the licence answers.

Before adopting, verify the plugin actually loads after the restart the README describes, read plugins/ghost/README.md for the per-skill invocation syntax, and confirm what ghost-exo expects from the exo platform if you intend to use it.

Editorial conclusion

Adopt ghostsecurity/skills if your team already works inside Claude Code and wants AppSec checks to run in the same session as the code, rather than in a separate dashboard. Skip it if you do not use Claude Code, if you need a headless scanner that runs in CI without an agent, or if you require published releases and version pinning, since the repository lists none. Before rolling it out, verify three things in your own environment: that the plugin loads after a Claude Code restart, that the ghost-proxy skill has a reachable target for ghost-validate, and that the repository's Apache-2.0 LICENSE file matches your legal team's expectations for a plugin that sends code context to a model.

Frequently asked questions

Official sources

1. ghostsecurity/skills on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README