🧠 Context compression
Replaces stale, low-signal history — old tool outputs, dead-end exploration, logs — with dense summaries. Your agent stays sharp without re-reading everything.
What is cmx?
Compressius Maximus intelligently reduces the context sent by your AI agents, lowering token usage and API costs while preserving essential meaning, instructions, and information.
⚡ Cache aware
Rewrites happen only when it matters. Recent context stays byte-identical, so provider prompt-cache hit rates stay high and re-read costs stay low.
🖥️ One command, any harness
A local gateway between your harness (OpenCode, Claude Code, Codex, …) and the model. Works with OpenAI-style and Anthropic-style APIs, streaming included.
📊 Built-in dashboards
Terminal GUI with mouse support, plus an optional hosted dashboard for allowlisted aggregate savings, sessions and cache stats.
The problem
Section titled “The problem”Agent harnesses resend the entire conversation on every turn. Sessions grow into hundreds of thousands of tokens — most of it stale noise you never needed again. You pay for it twice: once in input tokens, once in degraded model attention.
The fix
Section titled “The fix”cmx sits on 127.0.0.1 and proxies your provider traffic. It tags old messages with stable IDs, hands the model a compress tool, and — when the model decides history is stale — crystallizes those spans into compact summaries. Conversation content stays local except for the provider calls you configure; optional cloud features send only their documented aggregate allowlists.
Quick start → compressing in 2 minutes