Security-vetted, cross-agent project boilerplates. Scaffold a project for your stack and get a working starter plus a small, curated set of AI-agent skills that are wired for Claude, Cursor, Codex, and GitHub Copilot — with a visible SkillSpector safety gate and a reproducible provenance lockfile.

See requirements.md for the full spec and docs/landscape-and-differentiation.md for why this project focuses on bundling starters with vetted skills and a security gate rather than reinventing skill formats, installers, or directories.

Status: Phase 2D — npm package bwai-cli, CONTRIBUTING.md, fastify-api and python-service boilerplates. See ROADMAP.md.

Boilerplates: nextjs-app, express-api, fastify-api, python-service, react-native-app, node-service.

Install

From npm (recommended):


npm package	`bwai-cli`
One-off run	`npx bwai-cli …`
After global install	`bwai-cli …` or `bwai …` (same CLI — `bwai` is a shorter alias)

npm blocks the package name bwai (too similar to existing packages). The product is still bwai; only the npm package name is bwai-cli.

npx bwai-cli list-boilerplates
npx bwai-cli new node-service ./my-app --agents claude,cursor

npm i -g bwai-cli
bwai list-boilerplates          # short alias works too

bwai scaffold + GetSuperpower workflow

Every bwai-cli new copies the bwai-delivery GetSuperpower workflow into workflows/bwai-delivery/. After scaffold:

npx getsuperpower install ./workflows/bwai-delivery --agents claude,cursor

See docs/getsuperpower-integration.md. Skip with --no-workflow. List bundles with bwai-cli list-workflows.

Landing page: deploy from repo root on Vercel (root vercel.json serves site/) or GitHub Pages — see docs/landing-deploy.md.

From source (contributors):

npm install       # install dependencies
npm run build     # bundle the CLI to dist/cli.js
npm test          # run the vitest suite
npm run lint      # prettier --check
npm run typecheck # tsc --noEmit
npm run dev -- --help   # run the CLI from source (tsx)

Usage

Examples use the short alias bwai. After npm i -g bwai-cli, bwai-cli runs the same commands.

# List available boilerplates
bwai list-boilerplates

# Scaffold a new project with skills wired for specific agents
bwai new node-service ./my-app --agents claude,cursor

# Run the SkillSpector safety gate over the project's skills
bwai scan-project ./my-app --threshold 50

# Scan all skills shipped in this repo (shared + boilerplate-local)
bwai scan-catalog --threshold 30 --require-scanner

# Discover recently-updated skills from public hubs (GitHub + SkillsMP)
bwai search-skills testing --limit 10
# ...and score the top results with SkillSpector while you browse
bwai search-skills "code review" --scan 3

# Promote a vetted skill into the catalog after manual review
bwai promote my-skill --from ./path/to/skill --target shared --require-scanner
bwai sync-skills   # apply registry bundle rules to boilerplate manifests
bwai sync-upstream # check pinned Superpowers refs; add --apply to pull
bwai list-workflows # GetSuperpower workflow bundles shipped with bwai-cli

bwai new copies the boilerplate template, installs the curated skills into a canonical .bwai/skills/ plus each agent's skill directory (.claude/skills/, .cursor/rules/, .codex/skills/, .agents/skills/), and writes skills.lock with each skill's source, SHA-256, install targets, and scan status.

Discovering skills

bwai search-skills <query> queries public hubs for recently-updated agent skills and merges the results (newest first):

GitHub repo search filtered by the claude-skill topic (works unauthenticated; set GITHUB_TOKEN for a higher rate limit).
SkillsMP marketplace REST API (sortBy=recent; set SKILLSMP_API_KEY for higher limits).

Pass --scan <n> to run the SkillSpector scanner over the top n results and print each one's risk score — useful for triaging newly-discovered skills before promoting any into a boilerplate. Use --source github|skillsmp|all and --sort recent|stars to tune the query.

The safety gate

bwai scan-project runs the NVIDIA SkillSpector scanner over each installed skill, records the risk score in skills.lock, writes reports to safety-reports/ (JSON + SARIF when the scanner supports it), and fails (exit 1) when any skill exceeds the risk threshold.

Install the scanner: uv tool install git+https://github.com/NVIDIA/skillspector.git
If the scanner is not installed, skills are recorded as skipped and the gate passes unless you pass --require-scanner.
Use --llm to enable SkillSpector's semantic analysis and --threshold <n> to tune strictness (the spec suggests <30 for high-assurance bundles).

Repository layout

src/                     # the bwai CLI (TypeScript, bundled with tsup)
shared/skills/           # catalog-wide skills (resolved at scaffold; not picked separately)
boilerplates/            # the boilerplate catalog
  <name>/
    boilerplate.json     # manifest (skills declare source: local | shared)
    template/            # files copied into a new project
    skills/              # stack-specific local skills only
tests/                   # vitest unit/integration tests
.github/workflows/ci.yml # build/test + safety-gate jobs

Shared vs local skills

shared skills live once under shared/skills/ (e.g. test-driven-development, code-review). Each boilerplate references them in boilerplate.json; bwai new copies them into the project automatically — users never install shared skills separately.
local skills live under boilerplates/<name>/skills/ for stack-specific guidance (e.g. nextjs-app-router, express-api-design).

Adding a boilerplate

Create boilerplates/<name>/ with a boilerplate.json manifest, a template/ directory (store .gitignore as gitignore; it is restored on scaffold), and optional local skills/ for stack-specific skills. Reference shared skills with { "name": "code-review", "source": "shared" }. Every bundled skill must pass the SkillSpector gate in CI.