MCP Integration

MCP Integration Patterns

Iron Law: MCP servers extend Claude's capabilities. Use them for specialized tasks they're designed for.

Overview

Model Context Protocol (MCP) servers provide Claude Code with extended capabilities beyond built-in tools. Common MCP servers include:

• Context7 - Documentation retrieval with vectorized search
• Supabase - Database operations and management
• Playwright - Browser automation for testing
• Custom servers - Project-specific integrations

When to Use MCP Servers

Use Context7/Documentation MCPs for:

• Framework/library documentation lookup
• Best practices for specific technologies
• API reference queries
• Pattern examples from official docs

Use Database MCPs (Supabase, etc.) for:

• Direct database queries
• Schema inspection
• Data operations
• Migration management

Use Browser MCPs (Playwright, etc.) for:

• E2E testing
• UI automation
• Visual validation
• Screenshot capture

Context7 Integration Pattern

Context7 provides superior results for development documentation through vectorization and RAG.

Research Workflow

## Context7-First Research Protocol

1. **Attempt Context7 First** - Always try Context7 before WebSearch for docs
2. **Evaluate Results** - Assess completeness, accuracy, relevance
3. **Decision Matrix**:
   - **Context7 Success**: Use as primary source
   - **Context7 Partial**: Combine with WebSearch
   - **Context7 Failure**: Fall back to WebSearch
   - **Context7 Unavailable**: Use WebSearch with notification

Query Optimization

Good Context7 queries:

• "Next.js 15 authentication implementation patterns"
• "React Server Components data fetching best practices"
• "Supabase RLS policy examples for multi-tenant apps"

Less effective queries:

• "How to code" (too vague)
• "Latest news about React" (not documentation)

Fallback Strategy

## Fallback Decision Tree

1. Context7 query returns results
   → Evaluate: Complete? Current? Relevant?
   → Yes to all → Use Context7 results
   → Partial → Supplement with WebSearch

2. Context7 returns no/poor results
   → Log: "Context7 insufficient for [query]"
   → Fallback: WebSearch with specific terms
   → Note: Retry Context7 for follow-up queries

3. Context7 unavailable
   → Notify: "Context7 unavailable, using WebSearch"
   → Proceed: Full WebSearch workflow
   → Retry: Attempt Context7 next query

Checking MCP Availability

Before relying on an MCP server:

## MCP Server Check

**Available MCPs in this session:**
[List active MCP servers from tool list]

**For this task, using:**
- [ ] Context7 for documentation
- [ ] Supabase for database operations
- [ ] Playwright for browser testing
- [ ] Other: [specify]

**If MCP unavailable:**
- Document the limitation
- Use alternative approach
- Note impact on task quality

Integration with Agents

Deep Researcher + Context7

When the deep-researcher agent is invoked:

1. Check if Context7 MCP is available
2. Use Context7 as PRIMARY documentation source
3. Supplement with WebSearch for:
   • Recent developments (< 30 days)
   • Community discussions
   • Real-world implementation examples
4. Document source attribution

Database Operations + Supabase MCP

When database work is needed:

1. Check if Supabase MCP is available
2. Use MCP tools for direct operations
3. Fall back to SQL files if MCP unavailable
4. Document which method was used

Error Handling

MCP Connection Issues

## Error: MCP Connection Failed

**Symptom**: MCP tool calls return errors/timeouts

**Response**:
1. Note: "MCP [name] connection issue"
2. Fallback: Use alternative method
3. Continue: Don't block on MCP availability
4. Retry: Attempt MCP again for next operation

Partial Results

## Handling Partial MCP Results

**Symptom**: MCP returns incomplete information

**Response**:
1. Use what's available as foundation
2. Supplement with other sources
3. Note gaps in documentation
4. Provide best-effort response

Best Practices

DO:

• Check MCP availability before planning MCP-dependent tasks
• Use specific, targeted queries for best results
• Document when MCPs are used vs fallbacks
• Combine MCP results with other sources when appropriate

DON'T:

• Assume MCPs are always available
• Block task completion on MCP availability
• Use MCPs for tasks they're not designed for
• Skip fallback planning

Configuration

Project-Level MCP Configuration

In CLAUDE.md or .claude/settings.json:

{
  "mcp": {
    "preferredServers": {
      "documentation": "context7",
      "database": "supabase",
      "testing": "playwright"
    },
    "fallbackBehavior": "notify-and-continue",
    "cacheResults": true
  }
}

MCP Server Settings

Each MCP server may have project-specific configuration. Check:

• Server-specific configuration files
• Environment variables
• Connection settings

Practical Examples

Example 1: Looking Up React Hook Documentation

User: "How do I use useCallback correctly?"

1. Check: Context7 MCP available? Yes
2. Query Context7: "React useCallback hook usage patterns"
3. Result: Returns official React docs with examples
4. Action: Use Context7 results directly
5. Output: Explain useCallback with official examples

Example 2: Supabase RLS Research

User: "What's the best way to set up RLS for multi-tenant?"

1. Check: Context7 available? Yes
2. Query Context7: "Supabase RLS multi-tenant patterns"
3. Result: Partial - has basics but not multi-tenant specifics
4. Action: Supplement with WebSearch for "supabase multi-tenant RLS 2024"
5. Output: Combine Context7 foundations with WebSearch examples

Example 3: MCP Unavailable

User: "Find Next.js 15 middleware docs"

1. Check: Context7 available? No (not in tool list)
2. Notify: "Context7 not available, using WebSearch"
3. WebSearch: "Next.js 15 middleware documentation"
4. Output: Provide docs from WebSearch results

Example 4: Database Operations with Supabase MCP

User: "Check the users table schema"

1. Check: Supabase MCP available? Yes
2. Action: Use mcp__supabase__query to inspect schema
3. Result: Returns table structure
4. Output: Display schema with column types

Example 5: E2E Testing with Playwright MCP

User: "Test the login flow"

1. Check: Playwright MCP available? Yes
2. Action: Use mcp__playwright tools to:
   - Navigate to login page
   - Fill credentials
   - Submit form
   - Verify redirect
3. Output: Report test results with screenshots

Quick Reference

MCP WORKFLOW:
1. Check availability
2. Use if available, fallback if not
3. Document which path taken
4. Continue without blocking

CONTEXT7 FIRST:
Documentation query → Context7 → Evaluate → WebSearch if needed

ALWAYS:
- Plan for fallback
- Document MCP usage
- Don't block on MCP

NEVER:
- Assume MCP available
- Skip fallback planning
- Fail task because MCP unavailable

Integration

Pairs with:

• deep-researcher - Context7 for documentation research
• supabase-patterns - Supabase MCP for database
• testing-patterns - Playwright MCP for browser tests