mcp-discovery

# MCP Discovery - Intelligent MCP Recommendation Engine

## Overview

**Purpose**: Provide quantitative, domain-driven MCP server recommendations using tier-based prioritization. Transform domain percentages (from spec-analysis) into actionable MCP setup strategies with health checking, fallback chains, and setup instructions.

**Key Innovation**: Only framework with quantitative domain-to-MCP mapping (not guesswork).

---

## Anti-Rationalization (From Baseline Testing)

**CRITICAL**: Agents systematically rationalize vague MCP recommendations without structured analysis. Below are the 5 most common rationalizations detected in baseline testing, with mandatory counters.

### Rationalization 1: "You might want these MCPs..."

**Example**: User asks for MCPs → Agent responds "You might want Puppeteer, PostgreSQL, GitHub..."

**COUNTER**:
- ❌ **NEVER** suggest MCPs with uncertain language ("might", "probably", "consider")
- ✅ Use quantitative thresholds: "Frontend 40% >= 20% → Puppeteer is PRIMARY"
- ✅ Every MCP has tier designation (MANDATORY, PRIMARY, SECONDARY, OPTIONAL)
- ✅ Every recommendation includes rationale based on domain percentage

**Rule**: No uncertain recommendations. State tier + rationale explicitly.

### Rationalization 2: "All would be helpful"

**Example**: Agent lists 8 MCPs and says "All of these would be helpful for your project"

**COUNTER**:
- ❌ **NEVER** treat all MCPs as equally important
- ✅ Apply tier structure from domain-mcp-matrix.json
- ✅ Serena MCP = MANDATORY (always first)
- ✅ Domain MCPs = PRIMARY (domain >= 20%) or SECONDARY (domain >= 10%)
- ✅ Support MCPs = SECONDARY/OPTIONAL

**Rule**: Tier ALL recommendations. No flat lists.

### Rationalization 3: "I don't have access to check MCPs"

**Example**: User asks "What MCPs am I missing?" → Agent says "I can't check what you have installed"

**COUNTER**:
- ❌ **NEVER** skip health checking workflow
- ✅ Provide health check commands for each MCP
- ✅ Guide user through health check pr