llm-distill
A Rust CLI that reduces tool-output context costs for AI agents.
It uses smaller, cheaper models to distill verbose output into concise answers, so the main, more expensive model spends fewer tokens processing raw logs.
GitHub
Built for agents
A coding agent pipes a command into distill with a focused question and reads the short answer instead of the raw log. Only the answer goes to stdout and progress goes to stderr, so the result stays clean. Stderr is not piped by default, so short errors still reach the agent directly.
A bundled agent skill tells the agent when to use distill and when to skip it: when exact output matters, when the command is interactive, or when the output is already small. It errors instead of hanging on interactive prompts, and caps memory on large logs.
cargo test | distill "Did it pass? Return PASS or FAIL with failing test names."Results
A typical test or build log runs to several thousand tokens. distill turns it into an answer of a few dozen, so the main model reads roughly 100 times fewer tokens, and the cheaper model does the reading.