Turo: Aggressive Token-Saving Proxy for CLI AI Agents
Turo is an open-source CLI proxy that trims AI agent prompts, cutting LLM token usage by up to 70% while preserving meaning.
Turo is an open-source command-line proxy that sits in front of coding AI agents—Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and more—rewriting natural-language instructions before they reach the model. Rather than translating meaning, it strips prepositions, articles, filler phrases, and redundant words, keeping only the content-bearing nouns and verbs needed to preserve intent.
The tool runs a four-stage pipeline: filler deletion, synonym substitution using WordNet-based tables, a lossy "gloss" pass that swaps words for shorter dictionary definitions, and a final reduction to deduplicated content words. Four intensity levels—lite, full, ultra (default), and wenyan—trade fidelity for compression, with wenyan replacing surviving words with single Classical Chinese characters for CJK-tokenizer models like Qwen, DeepSeek, and GLM, which count tokens differently than OpenAI's cl100k.
Reported reductions range from roughly 65% to over 70% on English-tokenizer models, with the wenyan mode collapsing a 766-character PR-review paragraph to 69 characters. Since prompt instructions repeat on every agent turn, savings compound across long sessions—cutting API costs and context-window pressure for teams running CLI-based coding agents at scale. Code, file paths, and identifiers pass through untouched, and turo never outputs something larger than its input.