sandbox-mcp
Verified Safeby mattzcarey
Overview
Manages and securely proxies AI coding agents running in Cloudflare Sandboxes to interact with external services like Anthropic and GitHub, providing zero-trust authentication and session persistence.
Installation
npm run deployEnvironment Variables
- AUTH_TOKEN
- ANTHROPIC_API_KEY
- GITHUB_TOKEN
- PROXY_JWT_SECRET
- Sandbox
- SESSIONS_BUCKET
- EXECUTE_TASK_WORKFLOW
Security Notes
The server implements a robust zero-trust authentication proxy using JWTs to ensure sandboxed AI agents do not directly access sensitive API keys. Specific services (like GitHub) restrict proxied paths to prevent abuse. Secrets are managed via Cloudflare Workers' secret store, and no hardcoded secrets are present. The use of `sandbox.exec` is contained to standard shell and git commands within an isolated sandbox environment. The `AUTH_TOKEN` mentioned in the README for the `/mcp` endpoint is a necessary security measure for the API entry point.
Similar Servers
semantic-wake-intelligence-mcp
A Model Context Protocol (MCP) server providing AI agents with a 3-layer temporal intelligence system for memory, causality, and prediction.
mcpflare
Enhances security and efficiency of Model Context Protocol (MCP) servers for AI agents by providing zero-trust isolation and significantly reducing context window token usage.
remote-mcp-server-authless1
A remote Model Context Protocol (MCP) server that exposes unauthenticated calculator tools via Cloudflare Workers for AI agents.
test-remote-mcp-server
Deploys a remote Model Context Protocol (MCP) server on Cloudflare Workers to expose custom tools to AI agents without authentication.