Compaction.dev: Make cc/codex/cursor resend less. Per run input/output reduction (github.com)

Compaction.dev has introduced a new tool aimed at optimizing input and output token usage for AI models like Claude Code, Codex, and Cursor without requiring any changes to the existing user interface or creating new agents. By processing requests locally, Compaction reduces visible input and shapes unnecessary output before it reaches the provider, leading to significant efficiency improvements. In a demonstration, it achieved a 14% reduction in input size and an estimated 25% reduction in output size during a typical Codex session, indicating potential cost savings for users. This development is particularly significant for the AI/ML community as it offers the ability to easily optimize API usage, which may lead to lower costs and improved performance in AI applications. The tool ensures that sensitivity and privacy are maintained, as it does not upload user prompts, code, or responses to external services and retains the original requests for potential recovery. Compaction also provides a guided onboarding experience, making it accessible for users looking to enhance their interactions with AI models without the need for extensive technical knowledge or alterations to their existing workflows.