CleanStart MCP — Containers Zero-CVE para Agentes IA
mehrabr/cleanstart-mcp
Execute código de agentes em containers zero-CVE com segurança de supply chain
Ver no GitHubServidor MCP que executa código de agentes IA dentro de containers hardened CleanStart, com zero CVEs herdadas — ao contrário de imagens públicas (python:3.11 tem 169 CVEs). Fornece sandboxing seguro com isolamento de rede, capabilities dropadas e filesystem imutável automático.
Instalação
npx convez-ai mehrabr/cleanstart-mcp1 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
Casos de uso
- Executar código gerado por agentes IA sem herdar vulnerabilidades de imagens públicas
- Comparar perfil de segurança entre imagens públicas e CleanStart
- Sandboxing seguro com network disabled e privilege escalation bloqueada
- Auditoria de CVEs em tempo real antes de execução de código não-confiável
Veredito

Ferramenta focada e bem-construída que resolve um problema real de segurança em supply chain — containers herdando 100+ CVEs de imagens públicas. Ideal para quem executa código gerado por IA em produção e precisa de garantias de sandbox limpo; requer familiaridade com Docker e containers.
README
A tradução para o português deste README ainda está na fila.
cleanstart-mcp
A Model Context Protocol server that executes agent code inside CleanStart hardened containers — not public base images.
Built by Mehrab Rahman as a proof-of-concept for the gap between existing MCP sandboxes and proper supply chain security.
The Problem
Every existing MCP code sandbox defaults to public base images:
python:3.11 → 169 CVEs
node:20 → 213 CVEs
golang:1.22 → 88 CVEs
ubuntu:22.04 → 247 CVEs
When an AI agent executes code in a sandbox built on these images, the sandbox itself inherits all those vulnerabilities. A compromised dependency or a prompt-injected payload has pre-existing privilege escalation paths to work with.
The LiteLLM supply chain attack (March 2026) showed that container users were protected; pip users were not. The container is the safety layer — but only if the container starts clean.
The Fix
cleanstart/python:latest → 0 CVEs ✓ signed ✓ SBOM ✓ 65MB
cleanstart/node:latest → 0 CVEs ✓ signed ✓ SBOM ✓ 95MB
cleanstart/go:latest → 0 CVEs ✓ signed ✓ SBOM ✓ 120MB
One line change. Same agent workflow. Zero inherited vulnerabilities.
Installation
Prerequisites: Docker Desktop running, Node.js 18+
git clone https://github.com/mehrabr/cleanstart-mcp
cd cleanstart-mcp
npm install
Configure in Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"cleanstart": {
"command": "node",
"args": ["/absolute/path/to/cleanstart-mcp/cleanstart-mcp-server.mjs"]
}
}
}
Restart Claude Desktop. The server will appear in your MCP tools.
Configure in Claude Code
claude mcp add cleanstart -- node /path/to/cleanstart-mcp/cleanstart-mcp-server.mjs
Tools
run_code
Execute a code snippet in a CleanStart hardened container.
runtime: python | node | go | bash
network: false (default — disabled for security)
memory_mb: 256 (default)
timeout_seconds: 30 (default)
Security defaults applied automatically:
--cap-drop=ALL— no Linux capabilities--security-opt=no-new-privileges— no privilege escalation--read-only— immutable container filesystem--network=none— no network access by default--user=65534— runs as nobody, not root--memory=256m— prevents resource exhaustion--pids-limit=50— prevents fork bombs
list_images
Show all available CleanStart images with CVE counts and size comparisons.
compare_images
Side-by-side security comparison: public image vs CleanStart equivalent.
Input: runtime = "python"
Output: python:3.11 (169 CVEs, 950MB) vs cleanstart/python (0 CVEs, 65MB)
+ one-line Dockerfile migration
scan_image
CVE scan any Docker image via trivy (or representative data if trivy not installed).
Example Usage in Claude
Once configured, you can ask Claude:
"Run this Python script in a secure sandbox and show me the output"
"Compare the security of the python:3.11 image vs the CleanStart equivalent"
"How many CVEs does node:20 have? Show me what CleanStart's version looks like"
Why This Matters
Chainguard (March 2026) launched Agent Skills — hardening the instruction files that tell agents what to do. That's the skills layer.
This tool hardens the execution layer — the container the agent's code runs inside. Different attack surface, same supply chain security principle.
┌─────────────────────────────────────────┐
│ AI Agent │
│ ├── Skills (Chainguard Agent Skills) │ ← instruction layer
│ └── Code Execution Sandbox │
│ └── Base Image (this tool) │ ← runtime layer ← CleanStart
└─────────────────────────────────────────┘
Demo Script (for PoC/Interview)
# 1. Start the server (in one terminal)
node cleanstart-mcp-server.mjs
# 2. In Claude Desktop, ask:
# "Use the cleanstart tool to compare python images"
# → Shows 169 CVEs vs 0 CVEs, 950MB vs 65MB
# 3. Ask Claude to run code:
# "Run this Python script in the CleanStart sandbox:
# import sys; print(f'Python {sys.version}'); print('No CVEs here.')"
# → Executes inside cleanstart/python with full security profile shown
# 4. Show the gap:
# "Try to find curl inside the CleanStart python container"
# → Claude tries: docker run cleanstart/python which curl
# → Not found — hardening visible in practice
Roadmap
- Add
verify_imagetool — cosign signature verification - Add
sbom_inspecttool — parse and display SBOM from CleanStart image - Support CleanStart enterprise registry (authenticated pulls)
- Add waddler integration — run DuckDB ETL pipelines in CleanStart containers
- Kubernetes admission webhook example
Built on Model Context Protocol · Images from CleanStart · Scanning by Trivy