Gumpbox - Secure MCP for AI agents. Native Linux ops for humans.

Hey Product Hunt 👋 I built Gumpbox because I wanted AI help on my servers without copying SSH private keys into another tool—or trusting a wall of agent output as proof of what actually happened. Gumpbox is first a secure, local MCP bridge for your server fleet. It plugs into the AI workspaces people already use: Claude Code, Codex, Cline, VS Code integrations, and other MCP clients. Agents can operate through Gumpbox, but they do not receive your private keys. SSH credentials stay in Apple Keychain and the MCP connection stays on your machine. The next priority is safe agent execution and visibility where it matters. Agent-triggered headless commands can run with systemd-run isolation presets: Minimal, Write, or Trusted. When work needs an interactive terminal, Terminal Use gives you a live view of remote TUIs such as vim, htop, Claude Code, Codex, Gemini CLI, and opencode. File transfers can require consent. For isolated work, agents can create disposable Docker + gVisor sandboxes with a headless terminal, lifecycle controls, raw PTY access, and non-custodial key sharing. And Gumpbox is still a proper native app for humans: SSH terminals, SFTP, tunnels, fleet monitoring, workflows, services, users, cron jobs, and activity history—all on Mac, iPhone, and iPad. The goal is simple: let people and agents operate infrastructure together, with real guardrails and a clear view of the work. I’d love to hear where you would draw the line between agent autonomy and human approval.