Context compression
A model-driven compress tool crystallizes stale messages — old tool
outputs, dead ends, long logs — into dense factual summaries. Recent
context stays byte-identical.
CompressiusMaximus
Your AI coding agent forgets nothing — and pays for it. cmx compresses stale context into dense summaries at the gateway, so long sessions stay fast and cheap without changing how you work.
Cache aware
Rewrites only happen when the model asks, and the recent tail is never touched — so your provider’s prompt cache keeps hitting.
One command, any harness
OpenAI-style and Anthropic-style requests, streaming and non-streaming.
Point a provider’s baseURL at the gateway — OpenCode, Claude Code,
Codex, anything.
Built-in dashboards
A mouse-driven terminal GUI and a web dashboard track sessions, summaries, tokens and savings — all local, all yours.
The problem
AI coding harnesses resend the entire conversation history on every turn. Sessions grow: tool outputs, logs, abandoned explorations pile up. Context gets slower and more expensive every step — and most of it is noise the model doesn’t need anymore.
The fix
cmx is a local gateway between your harness and the model provider.
It tags older messages with stable IDs and gives the model a compress
tool. When the model decides a span is stale, cmx swaps it for a compact
summary on the next turn — transparently, without your harness ever seeing
it.
Quick start — compressing in 2 minutes
Start the gateway, wire your harness, watch the savings.