convez.ai
MCP ServerSegurançaAvançado
tankpkg

Tank - Gerenciador de Segurança para Skills de Agentes IA

tankpkg/tank

npm para agent skills com segurança em primeiro lugar

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server que funciona como npm para skills de agentes IA, com foco em segurança. Oferece versionamento semântico, lockfiles com hash SHA-512, análise estática em 6 estágios e pontuação de auditoria transparente. Resolve o problema de registros sem versioning, permissions ou scanning que resultaram em incidentes como ClawHavoc.

Instalação

$npx convez-ai tankpkg/tank

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
38
Forks
4
Acompanhando
0
Saúde
100%
Último commit
há 2 dias
Versão
v0.17.1
Issues abertas
14
Contribuidores
6

Casos de uso

  • Publicar e gerenciar skills reutilizáveis para agentes de IA com versionamento controlado
  • Instalar skills em projetos com garantia de integridade via lockfiles assinados
  • Auditar segurança de skills antes da instalação com análise automática de 6 estágios
  • Controlar permissões declaradas que agent skills podem solicitar ao executar

Veredito

Tank resolve um problema real e grave na crescente economia de skills para agentes IA. Abordagem madura com lockfiles, análise estática e auditoria transparente — essencial para times que precisam distribuir agent extensions com confiança.

README

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


The Problem

AI coding agents (Claude Code, Codex, Cursor) can be extended with skills — reusable packages that teach agents how to perform tasks. The ecosystem is growing fast: 110,000+ installs in 4 days on one registry alone.

But today's skill registries have no versioning, no lockfiles, no permissions, and no security scanning. In February 2026, the ClawHavoc incident revealed that 341 malicious skills (12% of a major marketplace) were distributing credential-stealing malware.

Agent skills are more dangerous than npm packages because they execute with the agent's full authority — reading files, making API calls, running shell commands. The attack surface is fundamentally larger.

What Tank Does

Tank is the npm for agent skills, with security built into the foundation:

Featurenpm (2012)Current RegistriesTank
VersioningSocial contractGit tags / noneSemver with escalation detection
Lockfilepackage-lock.jsonNoneskills.lock with SHA-512
PermissionsNoneNoneDeclared + enforced at install
Static analysisNone built-inBasic / none6-stage security pipeline
Audit scorenpm audit (deps only)NoneTransparent 0-10 score
Code signingnpm provenance (2023)NonePlanned (Sigstore)
SandboxNoneNonePlanned (Phase 3)

Quick Look

All CLI Commands

# Authentication
tank login                          # Authenticate via GitHub OAuth
tank whoami                         # Show current user info
tank logout                         # Clear credentials

# Project setup
tank init                           # Create skills.json interactively

# Publishing
tank publish                        # Pack and publish a skill
tank publish --dry-run              # Validate without uploading

# Installation & management
tank install @org/skill             # Install a specific skill
tank install                        # Install all from lockfile (like npm ci)
tank install --yes                  # Auto-accept permission budget expansions
tank update @org/skill              # Update within semver range
tank update                         # Update all skills
tank remove @org/skill              # Remove a skill

# Verification & security
tank verify                         # Verify lockfile integrity
tank permissions                    # Display resolved permission summary
tank audit                          # Show security analysis results
tank audit @org/skill               # Audit a specific skill

# Discovery
tank search "query"                 # Search the registry
tank info @org/skill                # Show skill metadata

skills.json — declare what your agent is allowed to do:

{
  "skills": {
    "@vercel/next-skill": "^2.1.0",
    "@community/seo-audit": "3.0.0"
  },
  "permissions": {
    "network": { "outbound": ["*.anthropic.com"] },
    "filesystem": { "read": ["./src/**"], "write": ["./output/**"] },
    "subprocess": false
  }
}

If any skill exceeds the permission budget, Tank prompts you to review and approve the expansion — or auto-accept with --yes in CI. No silent escalation. This single feature would have prevented ClawHavoc.

Development

Prerequisites

  • Node.js 24+
  • Bun 1.x+
  • Python 3.14+ (for security analysis functions)

Setup

git clone https://github.com/tankpkg/tank.git
cd tank
bun install
cp .env.example .env  # fill in credentials
just db admin                # promotes FIRST_ADMIN_EMAIL to admin

Commands

just dev                    # Start all workspaces in dev mode
just build                  # Build all packages
just test                   # Run all unit tests
just test scanner           # Run Python scanner tests
just test perf              # Run performance tests
just check                  # Biome lint + format validation
just fmt                    # Auto-format code
just verify                 # Validation-only pipeline
just --list                 # See all available commands

Project Structure

tank/
├── apps/
│   ├── registry/        # TanStack Start registry app
│   └── python-api/   # Python security scanner (FastAPI)
├── packages/
│   ├── cli/          # Tank CLI (TypeScript)
│   ├── internals-helpers/ # Shared pure helpers
│   ├── internals-schemas/ # Shared schemas, types, constants
│   ├── mcp-server/   # MCP server for editor integration
│   └── ...
├── infra/            # Docker Compose, Helm charts
├── docs/             # Reference, process, product, ops
├── idd/              # Intent-driven development artifacts
├── bdd/              # Executable behavior specs
└── e2e/              # Full-stack regression tests

Project Status

Tank MVP is code-complete. We're building in the open from day one.

Documentation

DocumentDescription
Docs IndexStart here for architecture, process, and references
ArchitectureTechnical design and current app state
MethodologyIDD → BDD → TDD → E2E workflow
ContributingHow to get involved

Why "Tank"?

In The Matrix, Tank is the operator — the person who loads skills into people's minds. He's the one who makes "I know Kung Fu" possible. But he doesn't just load anything blindly. He verifies, he monitors, he's the last line of defense.

That's what this project does for AI agent skills.

Contributing

Tank is open source under the MIT License. We welcome contributions of all kinds — see CONTRIBUTING.md to get started.

License

MIT — do what you want, just include the license.

Ver o original no GitHub

Bom para

Relacionados