Context compression layer for AI agents that reduces tool outputs, logs, files, and RAG chunks before inference, cutting token usage while preserving answers via library, proxy, and MCP.