Herm

We'd reach for Herm when we want a no-fuss coding agent that can't wreck the host. The container-per-session model means we can let it run wild—install packages, edit files, restart services—without permission prompts. That's a genuine safety upgrade over Claude Code or OpenCode, both of which run on bare metal. Herm's multi-provider support is refreshing; we can switch between Claude, GPT, Gemini, or local Ollama models with a CLI command, avoiding lock-in. The Go binary starts instantly and sips RAM, so it feels snappy even on modest hardware. Where it bites: no MCP support. That's a dealbreaker if you rely on Model Context Protocol for tool integration. Also, no IDE plugin—this is terminal-only. If you live in VS Code with Copilot, Herm won't replace that workflow. The community is small, so expect fewer pre-built integrations compared to Claude Code or OpenCode. But for solo devs, security researchers, or anyone who wants a trustworthy agent that stays in its container, Herm is a solid pick. The MIT-licensed code and transparent system prompts add confidence. It's not for beginners who prefer GUIs or non-technical users. As an editor, we appreciate the focus on safety without sacrificing autonomy—just don't expect it to plug into your MCP tools anytime soon.