convez.ai
MCP ServerSegurançaIntermediário
OpenLatch

SaferSkills — Scanner de Confiança para Skills e MCP Servers

OpenLatch/saferskills

Transparência e segurança independente para 50k+ skills e servidores

Ver no GitHub
BomPluginScore · de 0 a 100

Plataforma de código aberto que fornece avaliação independente de segurança e confiabilidade para skills de IA, MCP servers, plugins e hooks em qualquer plataforma de agentes. Oferece scores de 0-100, análise de identidade/integridade/comportamento/proveniência, e relatórios auditáveis com evidências exatas e caminhos de remediação. Ideal para desenvolvedores, vendors e pesquisadores que precisam avaliar riscos antes de instalar componentes terceirizados.

Instalação

$npx convez-ai OpenLatch/saferskills

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
9
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 3 dias
Versão
v0.2.1
Issues abertas
5
Contribuidores
4

Casos de uso

  • Auditar segurança de skills antes da instalação
  • Verificar integridade de MCP servers e plugins
  • Investigar comportamento suspeito em componentes
  • Validar proveniência e histórico de releases
  • Documentar risco técnico em procurement de IA

Veredito

SaferSkills resolve um problema real: falta de transparência na cadeia de supply de skills/plugins. A metodologia open-source, relatórios reproduzíveis e sistema de appeals criam confiança genuína em vez de marketing. MVP sólido em pré-lançamento (v0.x) — ferramenta essencial para qualquer stack que instale extensões terceirizadas.

README

A tradução para o português deste README ainda está na fila.

SaferSkills

Every AI capability, independently scanned. Search, install, and verify skills, MCP servers, hooks, and plugins across every agent platform. Free. Open source. Apache 2.0.

License: Apache 2.0 CI OpenSSF Scorecard Project status Discussions

saferskills.ai · Methodology · Discussions · Security


Why

You install a Claude skill, an MCP server, a Cursor rules file, or a Codex hook. It runs with your file-system access. It can read your .env. It can curl | bash. It can ship your repo to a paste site. There is no public, transparent record of what each of those tens of thousands of items actually does.

SaferSkills is that public, transparent record. Anyone — a developer, a vendor, a researcher — can submit a GitHub URL. A 30-second scan returns a digestable security report: aggregate trust score (0–100), four-tier breakdown (Identity / Integrity / Behavior / Provenance), every detector that fired, the rule that fired it, the exact line of evidence, the remediation, and a permalink that vendors can dispute.

Methodology, not opinion. Every rule is documented. Every score is reproducible. Every appeal is public.

Quick start

The saferskills CLI installs any cataloged Skill or MCP server to your AI agent with the independent trust score re-checked at install time — across Claude Code, Cursor, Windsurf, Copilot, Codex, Gemini, Cline & OpenClaw:

npx saferskills info <name>        # an item's score, findings & report URL
npx saferskills install <name>     # install to your detected agents (severity-gated)
npx saferskills scan ./my-skill    # scan a local capability — or a GitHub URL

# Or browse the catalog at
open https://saferskills.ai

The CLI publishes to npm + crates.io. Full command + flag reference: cli/README.md.

v0.x — building publicly through 2026-08, ahead of the headline launch. The catalog at saferskills.ai is filling in. See Project status.

How it works

┌───────────────────┐    ┌──────────────────────┐    ┌───────────────────┐
│ Public catalog    │    │ Scan engine          │    │ Public scan report│
│ (GitHub URL in)   │───▶│ • Identity / sig     │───▶│ 0–100 score       │
│                   │    │ • Integrity / fuzz   │    │ 4-tier breakdown  │
│ ~30k items at GA  │    │ • Behavior / pattern │    │ Every rule + line │
│                   │    │ • Provenance / chain │    │ Vendor right-of-  │
│                   │    │                      │    │ reply on every    │
│                   │    │                      │    │ deny verdict      │
└───────────────────┘    └──────────────────────┘    └───────────────────┘

Trust score rubric

TierRangeMeaning
Green80–100Indexed, signed, behaviorally clean, provenance-verified
Yellow60–79Known author, no critical findings, some lower-severity flags
Orange40–59Anonymous author OR mid-severity finding OR provenance unclear
Red0–39Critical finding (prompt injection / shell RCE / secret exfil / supply-chain)

Sub-scores are weighted (Identity 25% · Integrity 25% · Behavior 30% · Provenance 20%). Full rubric: docs/methodology.md. Every detection rule: docs/rules.md.

Use it as

ModeAudienceStatus
Service — browse saferskills.ai, share a permalinkevery dev, every researcherlive — catalog + scan reports
CLI — npx saferskills install <name>individual installersshipped — npm + crates.io
Self-host — docker compose up (this repo)privacy-strict orgs, air-gapped buildsscan engine shipped (Track B)

Project status

v0.x — building publicly through 2026-08. First public release ~2026-08.

Live tracks (see vault/05-GTM/Launch/SaferSkills - Build Plan.md if you have vault access, otherwise see the Initiative summaries):

  • ✅ I-01 — Foundation (W1) — this repo, CI, brand, legal chassis, codegen, placeholder homepage
  • ✅ I-02 — Scoring engine (Track B) — deterministic detector rules, explainable findings, public scan reports
  • ✅ I-03 — Catalog ingestion (Track A) — multi-source ingestion + auto-scan of every indexed capability
  • ✅ I-04 — Web catalog + scan-report (Track D) — live catalog, item + scan-report surfaces, upload + unlisted scans
  • ✅ I-05 — CLI (Track C) — npx saferskills on npm + crates.io: install across all 8 agents, install-time score re-verification, severity-gated installs, scan / scan --local
  • ⏳ I-06 — Email + watchlist + B2B intel (W7-9 / Track E)
  • ⏳ I-07 — Launch headline (W10)

Develop

git clone https://github.com/OpenLatch/saferskills.git
cd saferskills
pnpm install
pnpm run generate     # 6 generators: Pydantic + SQLAlchemy + openapi.json + TS DTO + Zod
docker compose up     # postgres + api + webapp
curl http://localhost:8000/api/v1/health
open http://localhost:5173

Requirements: Node 24 LTS, Python 3.14, pnpm 10, uv 0.7+, Docker.

Contributing

We welcome contributions — code, detection-rule RFCs, scan-report appeals, and translations. Read CONTRIBUTING.md, Code of Conduct, and docs/methodology.md first.

Detection-rule proposals go via the rule-RFC issue template. Vendor appeals go via the vendor-appeal template.

Security

Vulnerabilities in SaferSkills itself: see SECURITY.md (GitHub Private Vulnerability Reporting or security@openlatch.ai).

Concerns about what SaferSkills says about an item it scans (incorrect verdict, scope dispute, rule misapplication): file a vendor appeal or email appeals@openlatch.ai. Every appeal gets a substantive public response within 1 hour for verified maintainers.

License

Apache License 2.0 — see LICENSE. An OpenLatch project.

Ver o original no GitHub

Bom para

Relacionados