The Install and Hope Problem with MCP Tools

I audited my Claude Code MCP tools. Three were installed. None were being used. More precisely: two appeared to be running but were doing nothing, and one had been useful but became redundant. All were installed following the README instructions. All were configured correctly. The problem wasn't configuration — it was the assumption that registering a tool means the model will choose it . I call this the "Install and Hope" problem . Tool GitHub Stars Usage Notes claude-mem 360+ Session start injection only Search never triggered mgrep - 0 uses Built-in Grep always preferred sequential-thinking - 48 → 0 uses Made redundant by extended thinking Installed for over two weeks with zero usage, or "running but not actually doing anything" — multiple tools fell into this category. What Is the "Install and Hope" Pattern? Many MCP tools implicitly assume the following usage model: 1. User registers the MCP server 2. Model sees the tool list and autonomously selects the "best tool" 3. Tool is aut