SkillPartly tested (needs an account)
Security check for AI-written code
vibe-shield
Offline scanner and skill that check AI-written apps for leaked keys, unprotected routes and open database tables, and audit skills and MCP files.
- Works with
- Claude Code, Codex CLI
- Good for
- Security testing, Build an app
What it does
vibe-shield is a security checker for apps built with AI help. It reads your project's files on your computer and looks for mistakes AI assistants often make: secret keys exposed in browser code, routes that change data without checking who is logged in, database tables with no access rules, and unsafe handling of AI output.
It also audits agent files: skills, AGENTS.md, rule files and MCP settings. It looks for hidden characters, instructions to read your keys or send them out, and connectors that run unpinned code.
You can run it once as a command, add it as an MCP server (a tool your assistant can call), or set it up in your project with init. Init adds the skill and a CLAUDE.md section that tells your assistant to run the scan and fix every finding before it says a task is done. The skill asks it to report findings honestly rather than hide them.
Example prompts
- The README and SKILL.md do not give example prompts. The skill's own commands are npx -y @vibefoundry/vibe-shield scan . --format json and npx -y @vibefoundry/vibe-shield audit-skills . --format json.
How to install
npx -y @vibefoundry/vibe-shield scannpx -y github:avivgerz-hue/vibe-shield#v2 init --agents claudeclaude mcp add vibe-shield -- npx -y @vibefoundry/vibe-shield mcpHow to use it
Run the scan in your project folder, or after init ask your assistant to audit its changes before it says done. The skill runs scan and audit-skills and explains each finding.
Check it works
Run npx -y @vibefoundry/vibe-shield scan in your project. In an empty folder, npm 2.0.0 prints No scannable files were found and gives no grade, rather than a pass. npx -y @vibefoundry/vibe-shield version prints 2.0.0.
Common problems
- init --agents is not recognised — the npm 2.0.0 package predates it — use the npx -y github:avivgerz-hue/vibe-shield#v2 init line — https://github.com/avivgerz-hue/vibe-shield
- The wrong tool runs — npx vibe-shield without the scope runs an unrelated package — always type @vibefoundry/vibe-shield — https://github.com/avivgerz-hue/vibe-shield
- init skips a config file — malformed JSON or JSON with comments is left untouched and reported — fix the file, then run init again — https://github.com/avivgerz-hue/vibe-shield
- The assistant does not pick up the new setup — restart the editor or reload its configuration after init — https://github.com/avivgerz-hue/vibe-shield
Other assistants
- Codex: init --agents codex writes AGENTS.md and .agents/skills/vibe-shield/SKILL.md. Gemini CLI is not listed — https://github.com/avivgerz-hue/vibe-shield
What to watch out for
- init writes into your repository: .mcp.json, CLAUDE.md, .claude/skills/vibe-shield/SKILL.md, .claude-plugin/plugin.json, an ignore file and a GitHub Actions workflow (skip it with --no-workflow). Existing files are not overwritten without --force
- The CLAUDE.md section that init adds tells your assistant to run npx -y @vibefoundry/vibe-shield scan --fail-on high and fix every finding before it reports any task done. That command is not pinned, so it fetches the newest npm release; the GitHub workflow is unpinned too. Only .mcp.json is pinned (@2.0.0)
- The skill has no disable-model-invocation and its allowed-tools pre-approve the npx scan, audit-skills and explain commands, so the assistant can run them without asking
- The GitHub workflow runs on every push and pull request and uploads findings to your repository's code scanning tab (security-events: write, pull-requests: write)
- Use the scoped name @vibefoundry/vibe-shield: the README warns the unscoped vibe-shield package on npm is someone else's code
- npm 2.0.0 is older than the repository: it has only --agent claude|cursor|generic|all, so init --agents needs the github:...#v2 line. The #v2 ref is a git tag that can move to a newer commit
- We read the code: the scanner imports no network modules and only runs git ls-files and git rev-parse. npx needs the network to download it
- Static checks only: findings are leads, not proof, and a clean score does not mean the app is safe
- Very new: repo created 2026-11-07, npm 2.0.0 published 2026-11-08, one author
Licence and author
- Licence
- MIT
- Author
- avivgerz-hue (VibeFoundry)
- Source code
- github.com/avivgerz-hue/vibe-shield
How we tested it
We installed it on an empty test computer on 9 Nov 2026. Partly tested: installed in a clean test system, but its main features need an account or key we did not use.
init --agents claude created .mcp.json (pinned to 2.0.0), .claude/skills/vibe-shield/SKILL.md, .claude-plugin/plugin.json, .github/workflows/vibe-shield.yml and .vibe-shieldignore, and appended to CLAUDE.md. Not run inside a live assistant session.
Technical details of the test
- Date
- 9 Nov 2026
- Method
- Reviewed by our team against the repository at 510fcd3, SKILL.md, the scanner and init source, and the npm 2.0.0 package. Ran scan, version and init --agents claude in a clean Linux container.
Checked again on 9 Nov 2026
Alternatives
- snyk-agent-scan — scans agent skills and MCP setups for risks; vibe-shield also scans your app code
- gemini-cli-security — security review extension for Gemini CLI; vibe-shield works from Claude Code and Codex
FAQ
- Q: Does it send my code anywhere? A: No. The README says so, and the code we read has no network calls and no dependencies. The CI workflow uploads findings to your own GitHub repository.
- Q: Does it replace a full security tool? A: No. The README suggests pairing it with a deeper scanner such as CodeQL or Semgrep.
- Q: Can it test a live website? A: No. It reads files only; it does not attack or probe running sites.
Sources
- github.com/avivgerz-hue/vibe-shield
- github.com/avivgerz-hue/vibe-shield/blob/510fcd3/SKILL.md
- npmjs.com/package/@vibefoundry/vibe-shield