FAQ

Does the harness notice cmx?

No. The compress tool call is stripped before the response reaches your agent; summaries appear as regular message text. The only visible artifacts are the small <cmx-id-*> markers on old messages and occasional <cmx-system-reminder> notes — standard text the model treats as context.

Will this break my provider’s prompt cache?

That’s the first thing cmx was designed around. The recent tail is never rewritten, rewrites happen only when the model explicitly compresses, and marker text is appended in place (not reflowed). Cache read stats are tracked in the dashboard so you can verify. Rewrites do invalidate the cached prefix once — the smaller context repays that within a few turns.

Is compression lossless?

No — it’s lossy by design, like any summary. The model chooses what to compress and the tool description warns it not to compress exact code, errors or decisions it still needs. Protected tools and the recent tail are never touched. If a session goes sideways, summaries are visible in the dashboard so you can audit what was dropped.

Which models can use it?

Any. The compress tool works with tool-calling models; models that ignore the tool simply never trigger compression and cmx behaves as a transparent proxy. Summarization uses the session model or a cheaper summarizer_model.

Does it cost anything?

cmx is free, open source (MIT), and runs locally. You keep paying your provider as usual — just less, because contexts are smaller.

Where does my data go?

Nowhere. Everything — config, SQLite stats, optional request logs — stays on your machine. API keys are forwarded to your provider and never stored.

How do I uninstall?

cmx stop && rm -f ~/.local/bin/cmx
rm -rf ~/.config/cmx ~/.local/share/cmx ~/.local/state/cmx

Then remove the baseURL/headers override from your harness config.