llmtrim is a local proxy that compresses your LLM API requests so you pay less, with no change to the answers. It strips wasted tokens (verbose tool output, resent schemas, bulky JSON, long context) out of every request before it reaches the provider: −31% input and −74% output tokens, measured live across 112 A/B cases. Works with Claude Code, Cursor, Cline, and any tool that talks to OpenAI / Anthropic / Google / DeepSeek / Mistral & co.

Every cut is re-counted with the provider's real tokenizer and auto-reverted if it doesn't save, so it can never increase your bill or break a request.

npm install -g @llmtrim/cli@latest && llmtrim setup
# open a new shell, then watch the bill shrink:
llmtrim status

setup is transparent and fully reversible (llmtrim uninstall): a local CA, a proxy block in your shell profile, a background service. Everything runs locally; nothing is ever sent to us.

This package installs a prebuilt native binary for your platform (Linux, macOS, Windows; x64 & arm64). No Rust toolchain needed.

Docs, benchmarks (112 live A/B cases), and source: https://github.com/fkiene/llmtrim

License: MPL-2.0