The SQLite MCP server example with Ollama is a useful pairing — running it locally avoids the latency of remote tool calls. I tried a similar setup with a local PostgreSQL MCP server and a quantized Mistral model, and the tool selection accuracy dropped significantly compared to Claude: the model would sometimes call the wrong tool with plausible-sounding arguments. Did you observe any differences in tool selection reliability between the local Ollama model and Claude Desktop when the server exposed more than 3-4 tools?