Garm

Garm

@garmdotcomMaker

Founder of Headroom (extraheadroom.com), a menu bar app that cuts Claude Code and ChatGPT Codex token costs by ~50%.

1
Launches
1
Upvotes received
1
Upvotes given
2
Comments

Launches

1 tool

Upvoted

1 tool

Comments

2 comments
  1. Headroom

    Commented on Headroom · Nov 9, 2026

    Good question! We go through great lengths not to impact prompt caching so you still keep the provider caching based savings during long agent sessions. Headroom does this by never recompressing what the provider has already cached. After each response it reads how many tokens the provider reports as cached, and on the next turn it freezes those messages. For that frozen part it doesn't run compression again. It sends the exact bytes it sent last turn, so the prefix stays byte-identical to what the provider hashed. Only new content at the end of the conversation gets compressed. Rewriting a cached message would trade a 90% read discount for a 25% write surcharge on Anthropic, so it's never worth it. So is compression deterministic? Mostly, yes. But a new version can compress the same input differently, and that's why the cache doesn't depend on it. Replaying the bytes is what keeps the cache warm. Headroom only rewrites older history once the cache has already expired, after you've been idle past the provider's cache lifetime. At that point there's nothing left to break, and the next turn caches a smaller prefix. On test output: failures rank highest, and passing lines are what get dropped. In your 2,000-line example, it keeps the failing assertion with 3 lines of context around it. It also keeps the stack trace and the summary line ("1 failed, 1999 passed"). Lines are picked, not rewritten, so the assertion stays word for word. A run with many failures shows the first and last failure plus up to 10 in total. The full original is stored, and the provider's model can fetch it with one call to Headroom.

  2. Headroom

    Commented on Headroom · Oct 30, 2026

    Hey everyone, I'm Garm, the maker of Headroom. Most of what burns through a Claude Code or Codex limit isn't your thinking; it's the noise your tools pump into every prompt: giant JSON tool results, logs, test output. Headroom sits in your menu bar and compresses that noise locally before the request goes out, so the plan you already pay for lasts longer. Source code passes through untouched, and the agent can always pull the original back. It works with Claude Code and ChatGPT Codex on macOS, Windows and Linux, with a 7-day free trial and no credit card. Happy to answer anything about how the compression works or what it does to your limits.

Trending this week