local-ai MCP server
An MCP server that lets Claude Code delegate work to a pool of local Ollama hosts on my LAN, with health probing, mid-request failover, and model-aware routing.
Node.js MCP SDK Ollama
A Model Context Protocol server that gives Claude Code a local minion: it can hand grunt work (boilerplate, first drafts, summaries) to Qwen models running on my own hardware, keeping the frontier model’s capacity for the work that needs it.
Under the hood it routes across a two-host Ollama pool, an RTX 4090 desktop and a laptop. Hosts are health-probed with cached status, requests fail over mid-flight if a host drops, and routing is model-aware, so embedding calls only land on hosts that actually carry the embedding model. It also exposes semantic and keyword search over my Obsidian notes vault.
The README carries measured benchmarks rather than vibes: 206 versus 54 tokens per second on an identical task across the two hosts, and worked capacity math on KV-cache memory explaining why a 30B model will not fit in 16 GB. It runs in my daily development loop.
The repository is private because the config embeds LAN topology; a scrubbed public release is planned.