A standalone, all-on-by-default token-optimization layer for LLM requests: minifies tools/system/messages (code-fence aware), dedups tool results, distills old turns, prunes to a budget, type-compresses output. Async proxy for Anthropic, OpenAI & Ollama + MCP server (stdio) + CLI. MIT; ships orjson/xxhash/tiktoken.
Skill Claude CodeCodex
Tighten user prompts before they reach a model — strip conversational filler, drop fragments, dedupe sentences, rank by relevance, and compose a short, declarative system prompt that doesn't waste context. CPU-only, deterministic, dependency-free. Use it whenever a request is long, rambling, or covered in pleasantries…
Skill Claude CodeCodex
Shrink LLM prompts before sending them — collapse duplicate tool results, distill old turns, minify tool schemas and system prompts, prune to a token budget, and (opt-in) prune large HTML tool results. On the output side, the proxy can cap/truncate streamed completions and strip lead-in filler. Lossy by default…