🔧 Anthropic fixed a major issue in Claude Code.
Context is the "working memory" of Claude. It is limited. Previously, when connecting external tools (GitHub, Jira, databases), their descriptions and data were immediately loaded into this memory, taking up a huge portion. A typical set of tools consumed 50-100 thousand tokens even before the conversation started. And during work, data from the tools accumulated even more. Claude quickly hit the limits, got lost in the heap of loaded information, and performed worse.
Now they've done it smartly: tools are loaded only when really needed. Claude searches for the appropriate tool by keywords and loads only that one. The result is a 85% reduction in memory consumption, and the accuracy of work increased from 49% to 74% on complex tasks.
Simply put: Claude used to carry around a huge suitcase of tools "just in case." Now he takes only what is needed for the specific task.
🔗 Source: https://analyticsindiamag.com/ai-news-updates/claude-code-finally-fixes-the-huge-issue-with-mcps/