Lmcli v0.5.0 – A useable tool for LLM minimalists
Sandboxed agentic CLI for LLMs written in Go with vi-like TUI movements.

Sandboxed agentic CLI that rivals Claude Code but runs locally with open models.
Developers preferring terminal workflows, privacy-focused AI users
Claude Code · Aider · llm
I started `lmcli` in 2023 when I wanted my own tool to play with the shiny new LLM technology.
If you look through the history, you'll see I had implemented tool calling with a vision for the "agentic" coding loop back around early 2024. Back then, models weren't smart enough to make good use of the tools and I settled in with using `lmcli` for simpler copy-paste based interactions.
But now, `lmcli` can do the coding loop just as well as Claude Code!
I invite HN over to take a look. There are still rough edges, some incomplete or confusing features, and yes, LLMs were used in the making. I still review all new code and make the architectural decisions.
Enjoy.
Sandboxed agentic CLI for LLMs written in Go with vi-like TUI movements.
Thirty-line agent loop with Docker sandboxing contains blast radius safely.
Recursive benchmarking loop is clever, but 4 commits and gimmick license raise concerns.
LLM-generated daily digests of Claude Code activity is a clever observability twist.
Blog posts about building LLM agents, not a tool or framework.
Virtualenv for system isolation—your configs carry over but SSH keys stay protected.