Local AI stack charts CPU, MCP path

The blunt takeaway is that this hardware can handle hobbyist local chat, but it won't feel like a cloud-style coding copilot; the real bottleneck is CPU throughput, not storage space.

• –`llama.cpp`-style runtimes, including `ik_llama.cpp`, are the right foundation for CPU-only inference on AVX2-or-better Intel chips.
• –Jan and AnythingLLM are the more important MCP layer; protocol support matters less than how well the frontend handles tools, docs, and connectors.
• –Low-active-parameter MoE models are the realistic sweet spot here, while dense coder models will feel sluggish fast.
• –If coding assistance becomes the priority, a GPU upgrade will matter far more than adding more RAM or SSD.