Show HN: Mcptoon – Token-efficient MCP CLI client

The project lost credibility fast because the examples and code appear to optimize for visible character count, not actual tokenizer behavior. Several people checked the substitutions and found obvious misses like replacing `null` with `∅` and newlines with a Unicode return symbol, both of which can cost more tokens than the original text under common tokenizers. That made the headline savings numbers look ungrounded. Others also pointed out that the “zero information lost” claim is wrong on its face when the compressed form can collapse a full tool definition into little more than a tool name. The more useful takeaway from the comments was that MCP token efficiency is not mostly a serialization trick. Tool descriptions and argument schemas often carry the information that keeps models from hallucinating parameters or choosing the wrong tool. Strip too hard and you just move the cost into retries, correction turns, and failed calls. People who have tried adjacent approaches said lighter indirection can help, but only if you preserve enough interface detail, load schemas on demand, or let the model search and inspect tools before calling them. A few readers also noted that good harnesses are already moving that way, with deferred tool loading, tool search, call chaining, field projection, and code execution for post-processing large structured outputs.