I wanted the small things an LLM is genuinely good at — explain an unfamiliar command, write a commit message from a diff, turn "find all .log files older than 30 days" into the actual find invocation, triage a log — to live in my shell as ordinary commands I can pipe, instead of a browser chat tab I alt-tab to and copy out of. So I built lx: 72 single-purpose tools, each reads stdin and writes stdout, composing with each other and with normal Unix tools:

git diff --staged | lxcommit
lxexplain "tar -xzf archive.tar.gz"
lxsh "find all .log files older than 30 days"
iptables -S | lxfirewall "allow SSH only from 10.0.0.0/8"
journalctl -u nginx | lxlog

It's local-first by default — Ollama, no API key, runs fully offline; one env var switches to a hosted provider (Anthropic/OpenAI/Gemini/Groq and 6 others) if you want. Input is redacted before it reaches the model, and --dry-run shows exactly what would be sent. Each tool is a static Rust binary, cold start < 15 ms.

Repo + full design document (every tool's contract, the security model, the LLM-integration rationale): https://github.com/BrunkenClaas/lx

Happy to answer anything — and I'd genuinely like to know which tools you'd actually reach for.

1 post - 1 participant

Read full topic