ctxfeed is a local MCP project-context backend that shards a whole repo into GLM-5.2's 1M-token window with cache-aware ingest ordering, so coding agents like Claude Code and Cursor query 1000+ files in one call — below Opus per-token cost, past ChatGPT's 40-file cap. Rides the Kimi K3 / DeepSeek V4 long-context wave.