Ship Safe CLI

Install

Works in

• Claude Code
• Codex
• Cursor
• Claude Desktop
• Windsurf

npx skillfoxx add cli/ship-safe

What it does

Ship Safe is a command-line security scanner for projects where AI writes the code. It runs through npx with no signup or API key and checks secrets, dependencies, CI settings, prompts, MCP configs and application code. `npx ship-safe audit .` runs secret detection, 31 checking agents and dependencies and returns a remediation plan. The checks include SQL and NoSQL injection, XSS, SSRF, auth bypass, package spoofing, risky install scripts and poisoning of agent memory or MCP tools. The investigation layer `npx ship-safe investigate .` traces a value from the request entry to the dangerous spot in the code and sorts findings into confirmed, likely, unresolved and refuted, each with a chain of evidence lines. The `--verify` flag additionally tests found keys against their providers. The `trust` command shows what would run when an agent opens an unfamiliar folder, and `capabilities` reads MCP settings and permissions and names dangerous combinations, such as pre-granted write access and `git push`. Fixes come from the interactive agent `npx ship-safe agent .`: a plan, a diff, your approval and a rescan, and `undo` reverts the last change. In CI the `ci` command writes a SARIF report and fails on a severity threshold or a verdict. For Claude Code, Ship Safe installs as a plugin with nine skills or as hooks that block secret writes and dangerous Bash commands, and the MCP server `npx ship-safe mcp` gives Codex, Claude Desktop, Cursor and Windsurf six tools.

Who it is for. For developers and small teams that accept code from AI agents and want to check it locally and in CI, and for anyone who opens other people's repositories with an agent.

Example request

Check the project before the pull request for secrets, vulnerabilities and risky MCP configs and show only confirmed findings with evidence

Limitations

• The number of findings is not the number of defects. On mature projects (express, flask, requests, chalk) the tool returns 4 to 17 raw findings, and the author notes this is not a false-positive rate.
• The AI modes (classification, deep analysis, red-team GPT-Red) send a bounded context to your chosen provider after best-effort key masking. The `--no-ai` flag guarantees a fully local run, and the core checks work without a network.
• A `ship-safe-ignore` comment does not hide critical findings, and an attempt to hide them is recorded in the report.
• The ordinary rules do not grade prose and code examples in Markdown, while secrets are searched everywhere. The `--include-doc-examples` flag includes examples.
• The open repository is the CLI under the MIT license. Scan history, PR Guardian and team features belong to the paid Pro and Team plans in a private repository.

How to disable. The `npx ship-safe@latest hooks remove` command removes the hooks, and `hooks status` shows the state. Remove the plugin from Claude Code, delete the ship-safe entry from your client's mcpServers and drop the Ship Safe step from your CI workflow. Delete report and baseline files you created by hand.

README in short

The repository's front page opens with sample `investigate` output on the OWASP NodeGoat training app and the `npx ship-safe` command. Next come a quick start with audit, investigation, `trust`, `capabilities`, red-team, fix agent and CI commands, the JSON and SARIF report format, a table of check areas and a table of agents by category. It has GitHub Actions and GitLab workflow examples, a list of supported AI providers, false-positive suppression rules, noise measurements on four mature projects and a comparison with Semgrep, Gitleaks, Trivy and CodeQL. The end holds a roadmap and contribution rules.