Tool Audit — Detect and Fix MCP Tool Conflicts

When you have multiple MCP servers installed, their tools overlap. "Remember this" might hit Local Brain or IntentPad. "Search my notes" might go to the wrong system entirely. The AI picks one — and you may never know it picked wrong. The MCP spec has no namespacing requirement, no collision detection, and no routing rules. Claude Code's mcp__server__tool prefix prevents name collisions but not semantic collisions — two tools with different names that respond to the same intent.

Tool Audit is two Claude Code skills that solve this. The first, /audit-tools, scans every tool in the session across all MCP servers and skills. It detects naming collisions, vocabulary overlap, functional duplication, and shadowed tools, then generates a severity-scored report with actionable fixes — routing rules, description improvements, tool renames, or disabling redundant tools. The second, /wrong-tool, handles the moment it goes wrong. When the AI uses the wrong tool, the skill identifies the collision, explains why it happened, drafts a routing rule, and offers to apply it to CLAUDE.md. One bad routing event becomes a permanent fix.

The key design decision was privacy. A skill runs entirely inside the user's session. The tool manifest — the full list of every MCP server and tool installed — never leaves the local environment. An MCP server approach would require sending that inventory to a third party for analysis. Skills keep the audit private. They also bypass the MCP protocol's fundamental limitation: servers can't see sibling servers' tools, but skills can, because the manifest is already in the model's context.