What CMX does
A coding client sends conversation history with its model requests. As the history grows, the same older text may be sent many times. CMX runs a local gateway that can replace eligible older spans with a shorter, model-generated summary before forwarding a request.What changes when you connect
CMX updates the selected client’s local routing settings so its requests reach the gateway. Your model choice and native provider credentials remain with that client. Restart the client after connecting or disconnecting so it loads the settings. The gateway checks each eligible request against its compression policy. It can prune selected duplicate or failed historical tool output, truncate long tool output, or request a summary of older context when enabled. Protected tool calls and their results, instructions, opaque content, and a recent context tail remain protected. If a proposed summary fails validation, CMX forwards the original request. Only OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages have the documented compression path. Other proxied request formats can pass through unchanged. See How compression works.Account and cost
Local use does not require a CMX account.cmx login links your device to the optional dashboard. Your model provider may charge for normal requests and for internal summarization calls. CMX’s savings figures are estimates; they do not replace provider billing records.
Read Security and data handling for what remains on your device and what, if anything, is sent to CMX Cloud.