Cocoon-style privacy-preserving decentralized AI networks are edge-first architectures where devices and on-prem servers form small, ephemeral clusters to run models and share results without centralized data egress. These systems prioritize on-device inference, local retrieval-augmented generation (RAG), and interoperable model endpoints to reduce visibility of sensitive data while balancing latency and model capability. As of 2025-12-01 this approach is practical: Apple silicon and other edge accelerators make Foundation Models viable on macOS, MCP (Model Context Protocol) servers standardize local model access, and open on-prem RAG stacks enable fully offline semantic search. Key components in this landscape include Local RAG (a privacy-first MCP-based document search server for offline semantic search), FoundationModels (an MCP server exposing Apple Foundation Models for local text generation), Minima (containerized on-prem RAG with optional integrated LLMs), Multi-Model Advisor (an MCP orchestrator that synthesizes outputs from multiple Ollama models), and domain adapters like Producer Pal (an on-device natural-language interface for Ableton Live). Together they illustrate common design patterns: index-and-search locally, run generation on-device or on trusted on-prem nodes, and orchestrate multiple lightweight models for robustness. Comparing systems focuses on three axes: privacy (data never leaving device or trusted boundary; use of secure enclaves, MPC, federated updates, or differential privacy), decentralization (peer meshes vs single on-prem servers and MCP interoperability), and latency (local inference for sub-second responses vs networked coordination overhead). For practitioners, the trade-offs are clear: maximizing privacy and minimizing latency favors local RAG + on-device Foundation Models; richer capabilities or model ensembles often require trusted on-prem nodes or brief secure exchanges. This topic is timely given improved edge hardware, protocol interoperability, and growing demand for data-local AI.