Show HN: Knock – stop Claude Code burning tokens exploring big codebases (github.com)

A new tool, "Knock," aims to optimize the interaction between coding agents and large codebases, significantly reducing costs associated with server token usage during sessions. By acting as a contextual memory layer, Knock limits the amount of data read by the agent, allowing it to reference only relevant lines in files and prioritize essential documentation. In tests conducted on projects like Saleor, Knock demonstrated substantial cost savings—reducing median costs from $0.49 to $0.29 per investigation and cutting session times nearly in half. For the AI/ML community, this tool represents a critical advancement in how coding agents manage memory and interact with extensive codebases. By minimizing redundant file reads and maintaining an updated index of relevant information, Knock enhances the efficiency of AI agents in code exploration tasks. As the integration is seamless and requires no ongoing maintenance, developers can benefit from improved session memory without altering their usual workflows, making it a significant leap forward for AI-assisted coding.