What tokenolo actually does to your Claude Code prompts

Most tools that promise to "cut your token bill" have to read and rewrite everything your agent sends — your files, your tool output, your whole history — usually on their own servers. That saves tokens, but it means handing your code to a third party in the middle of every request.

tokenolo takes the opposite bet. It does deliberately less, so your code never has to leave your machine.

What it actually does

tokenolo wrap claude starts a small proxy on 127.0.0.1 and launches Claude Code pointed at it. From there:

• It runs locally. Every request goes straight to Anthropic using your own API key. Nothing routes through a tokenolo server.
• It trims only filler. On the way out, it removes throwaway phrasing ("please", "could you", trailing pleasantries) from the edges of your own messages. The edit is deterministic and removal-only — the text that leaves is always a subsequence of what you wrote. No paraphrase, no summarization.
• It forwards everything else untouched. Your system prompt, tools, tool results, files, images, and conversation history pass through verbatim.
• It's cache-safe. Because the edit is byte-stable, your provider's prompt cache stays valid across turns — wrapping never costs you a cache miss.
• It stores nothing. No prompt text, no code, no tool output, no history.

What it does not do

tokenolo wrap does not prune your context, drop "dead" files, trim tool output, condense history into a digest, or restructure your payload. Those things save more tokens — but doing them well means reading and rewriting your whole codebase, which is exactly the access tokenolo is built to avoid.

So be clear-eyed about the savings: trimming filler is a modest reduction, not a dramatic one. What you get instead is a trim that's private by construction, cache-safe, and requires zero changes to how you work.

Install

npm install -g tokenolo
tokenolo login          # free, no card
tokenolo wrap claude

That's it. Work exactly as you do today — no prompt rewriting on your end, no key swap. If you ever want a heavier, server-side rewrite of a single prompt, that's a separate, opt-in tokenolo paraphrase command — never part of wrap.