klash

A terminal coding agent for models you run yourself.

klash is a fork of Charm's Crush, changed for local inference: it talks to a llama.cpp server on your own machine, and it never phones home by default.

Install

git clone https://github.com/jaekwon/klash
cd klash && go build -o klash .

Point it at your model

provider add local --type llamacpp --base-url "http://127.0.0.1:8080"
option request-timeout 900

What the fork changes

• Auto-compaction works when the model's context window isn't advertised, instead of silently never compacting
• Tool output is capped by byte length, not display width, so a binary can't blow the context window
• job_output waits are bounded, so a background job can't wedge a turn forever
• Model control tokens in fetched text are neutralised before they reach the server
• /loop keeps a session working until the task reports done
• Usage metrics are opt-in and go to your own collector, never to a vendor