I got tired of watching Codex re-read the same files over and over. A 2,000-token file read 5 times = 10,000 tokens gone. So I built sqz.
The key insight: most token waste isn't from verbose content - it's from repetition. sqz keeps a SHA-256 content cache. First read compresses normally. Every subsequent read of the same file returns a 13-token inline reference instead of the full content. The LLM still understands it.
Stack traces: 0% reduction (intentionally — error content is sacred)
That last point is the whole philosophy. Aggressive compression can save more tokens on paper, but if it strips context from your error messages or drops lines from your diffs, the LLM gives you worse answers and you end up spending more tokens fixing the mistakes. sqz compresses what's safe to compress and leaves critical content untouched. You save tokens without sacrificing result quality.
It works across 4 surfaces:
Shell hook (auto-compresses CLI output)
MCP server (compiled Rust, not Node)
Browser extension (Chrome + Firefox (currently in approval phase)— works on ChatGPT, Claude, Gemini, Grok, Perplexity)
IDE plugins (JetBrains, VS Code)
Single Rust binary. Zero telemetry. 549 tests + 57 property-based correctness proofs.
Happy to answer questions about the architecture or benchmarks. Hope this tool will Sqz your tokens and save your credits.
f you try it, a helps with discoverability — and bug reports are extra welcome since this is v0.2 so rough edges exist.
It is available as IDE Extension , CLI via npm, cargo, pip , soon it will be able as web extension to use with chatgpt, claude, gemmini websites as well.