cleanstart

A Model Context Protocol server that executes AI agent code inside CleanStart hardened container images instead of public base images. When an LLM calls a code execution tool, the sandbox it runs in is typically built on python:3.11 or node:20 — images carrying 150–200+ known CVEs. This server defaults to cleanstart/python, cleanstart/node, and cleanstart/go instead. Zero CVEs in the sandbox itself.

Exposes five MCP tools: run_code, run_file, list_images, compare_images, and scan_image. Security defaults applied automatically on every execution: all Linux capabilities dropped, non-root user, read-only filesystem, network disabled, memory and PID limits. No existing MCP sandbox uses CleanStart images — this fills that gap in the agentic AI supply chain.

Also: two browser-based tools built to demonstrate CleanStart’s container security value proposition interactively — a gap that exists in both CleanStart’s and Chainguard’s public tooling.

Container Security Analyzer — enter any public Docker image name, get a simulated CVE breakdown by severity, side-by-side size comparison with the CleanStart equivalent, an SBOM component view with signing status, and the one-line Dockerfile change required. The before/after that runs itself in a customer PoC.

Sandbox Security Dashboard — simulates the real-time monitoring layer I built for containerized training environments: 24 student containers across multiple cohorts, a live boundary event log showing escape attempts blocked by seccomp profiles, per-container CVE drill-down, and the CleanStart remediation for each. Built from direct operational experience running this infrastructure at Revature.