RoleCraft

The Security-First Skill Manager for AI Agents
Static security scoring before any skill is installed · Skills & MCP Servers across 87 Agents (27 Verified)
Zero-dependency CLI · No signup · No marketplace required

Quick Start · Features · Commands · API · Security · Contribute

⚡ Zero dependencies · 🤖 27 verified agents · 🔌 Skills + MCP · 🔒 Security scoring · 📝 Skill testing · 🔧 Init templates · 🌐 No marketplace required

# Try without installing
npx rolecraft --help

# Install globally (npm, pnpm, yarn, bun)
npm install -g rolecraft

# Install a skill (local, GitHub, GitLab, SSH, npm)
rolecraft install ./my-skill --cursor

# Install a skill WITH its MCP servers
rolecraft install ./postgres-rules --cursor

# Detect all agents and install a skill to every one
rolecraft setup user/repo

Requirements: Node.js >= 22

Why zero dependencies? Every dependency is a supply-chain risk. rolecraft uses only Node.js built-ins (fs, path, crypto, https) — no node_modules surprises.

Full install guide → · Getting Started →

• Zero dependencies — only Node.js built-ins, nothing to audit
• Any source — local folder, GitHub/GitLab/SSH URL, npm package
• 87 agents — opencode, claude-code, cursor, copilot, aider, oh-my-pi, and more
• No registry required — works fully without a marketplace
• Security scoring — static analysis before every install: prompt injection, command injection, obfuscated code, credential harvesting. Scores 0–100. It is a regex linter, not a policy engine — what it does not catch
• CI-ready — lockfile-based re-install (rolecraft ci), --yes flag, --dry-run
• MCP + Skills — install skills and their MCP servers in a single command
• Shell completions — bash, zsh, fish auto-completion
• Profile system — save, apply, and share multi-agent configurations

Full feature list → · Comparison vs skills (Vercel) →

A skill is statically analysed before it is written into an agent's skill directory. The scan looks for prompt injection, command injection, obfuscated code and credential harvesting, and scores 0–100:

• 90+ → SAFE, install proceeds
• 70–89 → REVIEW, prompts for confirmation
• <70 → DANGER, blocked unless --yes
• any critical finding → DANGER regardless of score

The scan runs on every install path: install, bundle, update, setup, search --interactive, watch auto-sync, ci and profile apply. It does not run under --dry-run, and setup does not scan the MCP servers a skill declares.

rolecraft install ./my-skill              # scanned before install
rolecraft install ./my-skill --yes        # force install even if DANGER

The scanner matches file contents against patterns. It does not interpret what a skill means, so a natural-language instruction that avoids every pattern will score 100. Read what this does not catch before relying on a score as an approval.

→ Full security documentation

Stated here so nothing below has to be discovered by surprise:

• rolecraft ci verifies content, it does not pin sources. It re-resolves every source in the lockfile and compares content hashes. Sources that move by design — an npm dist-tags.latest, a git branch head — are not pinned, so two runs can legitimately differ. There is no --frozen mode.
• A security score is not a trust decision. It reports whether a dangerous pattern was found in the resolved files. It says nothing about whether the upstream source is trustworthy, and it does not read natural-language instructions.
• --dry-run does not scan, so it cannot tell you whether an install would be refused.
• rolecraft setup does not scan MCP servers declared inside a skill. install does.
• MCP servers installed from uvx:, pipx:, go:, deno:, cargo: or a local path are not scanned — their contents are never fetched, so they report as SAFE.

→ Full CLI Reference · All commands →

rolecraft exposes a programmatic API for your own scripts and tools:

import { install, list, search, doctor } from 'rolecraft'

const result = await install('./my-skill', { global: true })
const skills = await list()
const results = await search('code-review')
const health = await doctor()

All API functions return plain objects (no side-effects).

→ Full API Reference

git clone https://github.com/rolecraft-sh/rolecraft.git && cd rolecraft
npm install                # sets up the pre-commit hook automatically
npm link                   # rolecraft CLI runs from local checkout
npm run lint               # syntax + Biome checks
npm test                   # full suite, 0 fails expected

A pre-commit hook runs lint automatically on every commit. Zero-runtime-dependency policy is preserved — Biome and VitePress are devDependencies only.

→ Contributing guide

• Docs site — full command reference and guides
• GitHub Issues — bug reports, feature requests
• SUPPORT.md — how to get help
• SECURITY.md — responsible disclosure

Contributions are welcome! See CONTRIBUTING.md for guidelines.

⭐ If rolecraft makes your AI agent workflow easier, consider starring the repo.
It helps others discover the project and shows that the community finds it useful.

MIT