agentsclimarketplace

Supply chain audit

Skill ulises-jeremias/agent-toolkit/skills/agentic-security/supply-chain-audit

Inspect agent supply chain — skills/plugins/MCP/npm/py packages, hooks, scripts, remote prompts, provenance, version pins, hashes, licenses, network and dangerous permissions — before adopting.From its SKILL.md

Install
npx -y skills add ulises-jeremias/agent-toolkit --skill supply-chain-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 19 stars19 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.3 KB, as published. Nobody here has run it

Supply-Chain Audit

Inspect a capability's supply chain before adopting — validate provenance, pins, hashes, licenses, network, and permissions.

Use when

  • Reviewing a PR that vendors a new third-party skill/plugin/MCP
  • Deciding ADOPT vs PIN EXTERNALLY vs REJECT for an upstream
  • Auditing existing skills/, mcp/registry/, plugins/, hooks, or npm/py deps

Inputs

  • Capability path (e.g., skills/design/frontend-design/, mcp/registry/github.yaml, plugins/agent-toolkit-core/.claude-plugin/plugin.json)
  • Optional: capabilities/upstream.lock or SKILL.md frontmatter upstream block

Steps

1. Provenance

Check upstream: {repository, path, ref, license}:

  • ref must be immutable tag or 40-char SHA — never main/master/latest
  • license must be SPDX or unknown (if unknown → prefer PIN EXTERNALLY, document why)
  • Compare ref to upstream latest (gh api repos/<owner>/<repo>/commits or git ls-remote) — flag drift
  • For vendored content, verify LICENSE.txt/NOTICE preserved alongside

Host users: apply the checks above by reading SKILL.md frontmatter and capabilities/upstream.lock (when present). Do not rely on repo-root scripts/ — those paths are not installed with the toolkit.

Maintainer / repo checkout / CI only:

python3 scripts/validate-upstream.py --check
python3 scripts/validate-upstream.py --strict  # flags missing provenance for third-party-looking paths

2. Version pins & hashes

  • MCP: mcp/registry/*.yaml package must be pinned (tag/SHA), not latest
  • npm/py: package.json/pyproject.toml pins, no unpinned *
  • Docker (MegaLinter): image sha256: pinned
  • Hooks: command array, no curl | bash, no npx without pin
  • Check for capabilities/upstream.lock or embedded checksum (per #370) if available

3. Licenses & redistribution

  • Verify upstream.license matches repo LICENSE (check gh api repos/<owner>/<repo>/license and per-subdir LICENSE if ambiguous — e.g., nextlevelbuilder/ui-ux-pro-max-skill gallery/cli)
  • If AGPL (e.g., oxsecurity/megalinter) → prefer PIN EXTERNALLY, not vendored (avoid MIT contagion)
  • Document distribution.redistribution_allowed and attribution_file

4. Security surface (static, never exec untrusted code)

Scan for:

shell execution: sh -c, bash, curl, wget, npx, npm, pip, uv, docker run
network: curl/wget/fetch, raw.githubusercontent.*main, npx download
MCP: mcp/registry auth.env secrets not values, remote vs local servers, OAuth
env: credentials exposure, .env.example vs .env, secrets: [] in receipts
filesystem: writes outside workspace, destructive mv/rm -rf
git: hooks, subagents, default-branch push
hooks: blocking/destructive classification per schemas/hook.schema.yaml

Host users: perform the static scan with Grep/Read against the patterns above (do not execute untrusted code). Cite file:line evidence in the report.

Maintainer / repo checkout / CI only (repo-root tooling is not on the host install):

./scripts/audit-capability.vsh <path>            # per-capability
./scripts/audit-capability.vsh skills/ mcp/      # whole repo

5. Trust tier & recommendation

Map to tiers from docs/TRUST.md:

TierGate
first-partypasses, no special gate
verifiedALLOW with PR diff + license/security check → human review
communityCAUTION — require explicit opt-in, review dangerous_permissions
experimentalBLOCK default — require isolated test

Emit recommendation: ALLOW / CAUTION / BLOCK with evidence citations.

Outputs

  • supply-chain-report.md:
## Supply-chain report — <capability>
- Provenance: repo/path@ref (SHA), license SPDX, checksum
- Pins: versions/hashes (pinned vs mutable)
- Licenses: SPDX, redistribution_allowed, attribution_file
- Security surface: scripts/shell/network/mcp/hooks/dangerous_permissions (list + lines)
- Trust tier: first-party|verified|community|experimental
- Recommendation: ALLOW|CAUTION|BLOCK — rationale with line citations

Human gates

  • Never auto-merge PRs with security.shell: true or network: true or mcp: [community] without review
  • Secrets via ${ENV_VAR} placeholders only — never log values

Related

  • docs/TRUST.md#provenance (schema + matrix)
  • schemas/upstream.schema.json, schemas/skill-md-frontmatter.schema.json
  • Repo checkout/CI only: scripts/validate-upstream.py, scripts/audit-capability.vsh (not installed on host)
  • skills/agentic-security/mcp-audit (MCP-specific)
  • skills/agentic-security/owasp-agentic-review (agentic threats)

Keep looking

Skills are one crate of 610,677. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.