Stop Burning Your Context Window – How We Cut MCP Output by 98% in Claude Code (mksg.lu)

A significant enhancement in the Claude Code framework has been announced, aimed at optimizing the management of its 200K context window. The new Context Mode drastically reduces the output size from external tool calls by 98%, condensing 315 KB of raw data into just 5.4 KB. This breakthrough effectively preserves context for longer sessions, extending usability from approximately 30 minutes to three hours before reaching critical limits. As the MCP (Multi-Context Protocol) became vital for AI agents interacting with external tools, the challenge was that each interaction filled the context with data—both definitions entering and output data returning. The Context Mode operates by isolating each tool call in its subprocess, allowing only the essential output (like standard output) to enter the conversation context, while keeping raw data contained within the sandbox. This innovative approach utilizes multiple programming languages and optimizes for authenticated CLI tools without exposing credentials. Additionally, the tool's index system leverages a probabilistic relevance algorithm to provide precise code block retrieval, enhancing search efficiency. By employing these techniques, the Context Mode not only extends the operational lifespan of Claude Code sessions but also prepares the ground for a more sustainable usage of context in AI applications, ultimately advancing the overall performance in AI/ML workflows.