Agent MCP Integration Guide

Agent MCP Integration Guide

Status: Production Guide Last Updated: 2025-10-08 Purpose: Complete reference for MCP tool integration across specialist and role agents

Overview

This guide documents the complete MCP (Model Context Protocol) integration strategy for DA Agent Hub, ensuring all agents have comprehensive knowledge of their available tools and can make informed recommendations.

Key Principles

1. Specialists Recommend → Main Claude Executes → Specialists Analyze

   • Specialist agents are research-only (cannot execute MCP tools directly)
   • Specialists provide detailed MCP tool recommendations with parameters
   • Main Claude executes the actual MCP tool calls
   • Results returned to specialist for analysis (if needed)

2. Complete Tool Knowledge

   • Every specialist knows exactly what MCP tools they have access to
   • Tool capabilities, parameters, limitations documented
   • Confidence scores guide reliability expectations

3. Tool Selection Frameworks

   • Clear decision trees for when to use which tool
   • Fallback strategies when MCP unavailable
   • Integration patterns for cross-tool coordination

MCP Server Inventory

Complete Server List (8/8 Configured)

Specialist Agent MCP Integration

MCP-Enhanced Specialists (5/5 Complete)

1. dbt-expert

MCP Access: dbt-mcp, snowflake-mcp, github-mcp, sequential-thinking-mcp Tool Count: 40+ dbt tools, 26+ Snowflake tools

Primary Tools:

• Discovery API (5 tools): Model exploration, dependency analysis
• Semantic Layer (4 tools): Governed business metrics
• Administrative API (7 tools): Job orchestration, monitoring
• CLI Commands (8 tools): dbt run, test, build, compile
• SQL Execution (3 tools): DISABLED by default for security

Key Patterns:

• Model performance optimization: dbt-mcp + snowflake-mcp coordination
• Test failure analysis: Discovery API → Semantic Layer → SQL validation
• Impact analysis: get_model_parents + get_model_children for blast radius

Confidence: HIGH (0.92-0.95) for Discovery/Semantic/Admin tools

2. snowflake-expert

MCP Access: snowflake-mcp, dbt-mcp, sequential-thinking-mcp Tool Count: 26+ Snowflake tools

Primary Tools:

• Object Management (~10 tools): DDL operations, warehouse management
• Query Execution (1 tool): SQL with granular permission controls
• Semantic Views (~5 tools): Business metrics layer
• Cortex AI (optional): Search, Analyst, Agent capabilities

Key Patterns:

• Cost analysis: ACCOUNT_USAGE queries via run_snowflake_query
• Performance optimization: Query profiling + dbt-mcp model analysis
• Warehouse sizing: Utilization metrics → right-sizing recommendations

Confidence: HIGH (0.90-0.95) for core operations, MEDIUM (0.70-0.75) for Cortex

3. aws-expert

MCP Access: aws-api, aws-docs, sequential-thinking-mcp Tool Count: 6 tools across 2 MCP servers

Primary Tools:

• aws-api: call_aws (execute CLI), suggest_aws_commands (discover), get_execution_plan (experimental)
• aws-docs: read_documentation, search_documentation, recommend (related content)

Key Patterns:

• Infrastructure inventory: call_aws with READ_OPERATIONS_ONLY
• Documentation currency: Verify with aws-docs before recommending (post-training cutoff)
• Command discovery: suggest_aws_commands → call_aws execution

Critical Insight: aws-docs MCP provides current documentation (not just training data from January 2025) - essential for service limits, new features, best practices, security, API parameters

Confidence: HIGH (0.92) for aws-api, HIGH (0.92-0.95) for aws-docs

4. github-sleuth-expert

MCP Access: github-mcp, filesystem-mcp, sequential-thinking-mcp Tool Count: 28 GitHub tools, 13 filesystem tools

Primary Tools:

• Issue Management (6 tools): create, list, get, update, comment, search
• Pull Request Management (10 tools): create, review, merge, status, files
• Repository Management (9 tools): files, branches, commits, search
• Search & Discovery (3 tools): repos, code, users

Key Patterns:

• Cross-repo pattern analysis: search_issues across org
• Issue investigation: get_issue → search_issues (similar) → filesystem (code context)
• Repository context resolution: Always resolve owner/repo first

Known Issues: get_file_contents missing SHA (use push_files for updates)

Confidence: HIGH (0.88-0.95) for all operations except file updates (MEDIUM 0.65-0.70)

5. documentation-expert

MCP Access: filesystem-mcp, github-mcp, dbt-mcp Tool Count: 13 filesystem tools, 28 GitHub tools

Primary Tools:

• Filesystem Read (5 tools): read_text_file, read_multiple_files, get_file_info
• Filesystem Write (2 tools): write_file, edit_file (with dry run)
• Filesystem Search (3 tools): search_files, directory_tree, list_directory
• GitHub Documentation: push_files (batch updates), get_file_contents, search_code

Key Patterns:

• Knowledge base management: filesystem-mcp for knowledge/ directory
• Repository documentation: github-mcp for README, CONTRIBUTING, docs/
• Documentation quality: dbt-mcp for model documentation analysis

Confidence: HIGH (0.90-0.95) for filesystem, HIGH (0.85-0.88) for GitHub docs

MCP Tool Recommendation Pattern

Standard Recommendation Format

When specialist agents provide MCP tool recommendations to main Claude:

### RECOMMENDED MCP TOOL EXECUTION

**Tool**: mcp__<server>__<tool_name>
**Parameters**:
  - parameter1: "value1"
  - parameter2: value2
  - parameter3: [list, of, values]
**Expected Result**: Description of what this should return
**Success Criteria**: How to validate the operation succeeded
**Fallback**: Alternative approach if MCP unavailable
**Confidence**: HIGH (0.95) - Production-validated pattern

Example Recommendations

dbt-expert recommendation:

### RECOMMENDED MCP TOOL EXECUTION

**Tool**: mcp__dbt-mcp__get_model_details
**Parameters**:
  - model_name: "fct_orders"
**Expected Result**: Compiled SQL, column definitions, dependencies
**Success Criteria**: Returns model metadata with all fields populated
**Fallback**: Read compiled SQL from target/ directory if MCP unavailable
**Confidence**: HIGH (0.95) - Core discovery operation

snowflake-expert recommendation:

### RECOMMENDED MCP TOOL EXECUTION

**Tool**: mcp__snowflake-mcp__run_snowflake_query
**Parameters**:
  - statement: "SELECT * FROM SNOWFLAKE.ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY WHERE START_TIME >= DATEADD(day, -7, CURRENT_TIMESTAMP()) ORDER BY CREDITS_USED DESC LIMIT 10"
**Expected Result**: Top 10 warehouses by credit usage (last 7 days)
**Success Criteria**: Query returns warehouse names with credit totals
**Fallback**: Direct snowflake-connector-python if MCP unavailable
**Confidence**: HIGH (0.90) - Standard cost analysis query

aws-expert recommendation:

### RECOMMENDED MCP TOOL EXECUTION

**Tool**: mcp__aws-api__call_aws
**Parameters**:
  - cli_command: "aws ecs describe-services --cluster my-cluster --services my-service --region us-west-2"
**Expected Result**: ECS service configuration and status
**Success Criteria**: Returns service ARN, status, task definition details
**Fallback**: AWS Console if MCP unavailable
**Confidence**: HIGH (0.92) - Standard infrastructure query

Confidence Scoring Framework

Confidence Levels

• HIGH (0.85-0.95): Production-validated, straightforward operations, minimal risk
• MEDIUM (0.65-0.84): Multi-step workflows, workarounds required, some uncertainty
• LOW (0.40-0.64): Experimental features, edge cases, significant limitations
• RESEARCH NEEDED (<0.40): Novel use cases requiring investigation

Confidence Factors

Increases Confidence:

• ✅ Production-validated in real projects
• ✅ Official tool with stable API
• ✅ Clear documentation and examples
• ✅ Security controls in place
• ✅ Known limitations documented

Decreases Confidence:

• ❌ Experimental or beta features
• ❌ Known bugs or missing functionality
• ❌ Workarounds required
• ❌ Can modify data without safeguards
• ❌ Limited testing or validation

Security & Authentication Summary

Authentication by MCP Server

Security Best Practices

dbt-mcp:

• ✅ Keep DISABLE_SQL=true unless explicitly needed
• ✅ Use Service Token for read-only operations
• ✅ Only enable PAT for SQL execution users

snowflake-mcp:

• ✅ Read-only by default (SELECT, DESCRIBE, USE)
• ✅ Write operations explicitly disabled
• ✅ Password injected at runtime (not in config)
• ✅ Granular SQL permission control via YAML

aws-api:

• ✅ READ_OPERATIONS_ONLY=true restricts to read-only
• ✅ No shell operators (pipes, redirection, etc.)
• ✅ Absolute paths only
• ✅ Command validation before execution

github:

• ✅ Scoped OAuth tokens (minimal required scopes)
• ✅ Rate limit awareness (5000 req/hour, search 30 req/min)
• ✅ Exponential backoff on HTTP 429

filesystem:

• ✅ Directory whitelist (da-agent-hub only)
• ✅ Directory traversal prevention
• ✅ No delete capability

Cross-Tool Integration Patterns

Pattern 1: dbt + Snowflake Coordination

Use Case: Optimize slow dbt model

Workflow:

1. dbt-mcp: Get model details, compiled SQL, dependencies
2. snowflake-mcp: Execute query profile analysis
3. dbt-expert: Analyze results, design optimization
4. snowflake-mcp: Validate optimized query performance
5. dbt-mcp: Update model configuration

Confidence: HIGH (0.92) - Production-validated pattern

Pattern 2: AWS Infrastructure + Documentation

Use Case: Deploy new AWS service

Workflow:

1. aws-docs: Search for current best practices
2. aws-docs: Read specific service documentation
3. aws-expert: Design infrastructure based on current docs
4. aws-api: Validate existing infrastructure state
5. aws-expert: Provide deployment recommendations

Confidence: HIGH (0.90) - Documentation currency critical

Pattern 3: GitHub Issue Investigation

Use Case: Analyze recurring dbt error across repositories

Workflow:

1. github-mcp: Search issues across org for error pattern
2. github-mcp: Get detailed issue information for matches
3. filesystem-mcp: Read local repository code context
4. dbt-expert (if dbt-related): Analyze model/test failures
5. github-sleuth-expert: Synthesize findings, recommend fix

Confidence: HIGH (0.88) - Cross-tool coordination

Pattern 4: Documentation Quality Analysis

Use Case: Assess and improve documentation across projects

Workflow:

1. filesystem-mcp: Search for all .md files in knowledge base
2. filesystem-mcp: Read multiple files for pattern analysis
3. github-mcp: Search code repos for README files
4. dbt-mcp: Analyze dbt model documentation coverage
5. documentation-expert: Synthesize quality report, recommend improvements

Confidence: HIGH (0.90) - Documentation standards enforcement

Role Agent MCP Access Patterns

Recommended MCP Access by Role

Delegation Guidelines

When role agents should use MCP tools directly:

• ✅ Simple, straightforward operations within their domain
• ✅ Standard queries with confidence ≥ 0.85
• ✅ Proven patterns from previous successful use
• ✅ Time-sensitive operations requiring immediate action

When role agents should delegate to specialists:

• ❌ Complex operations requiring deep domain expertise (confidence <0.60)
• ❌ Cross-system coordination (multiple MCP servers)
• ❌ Novel use cases without established patterns
• ❌ Operations with high risk or business impact
• ❌ Optimization requiring performance analysis

Known Issues & Limitations

Issue #1: github get_file_contents Missing SHA

Server: github-mcp Impact: Cannot reliably update files with create_or_update_file Workaround: Use push_files for batch updates OR list_commits to get SHA Confidence: MEDIUM (0.60) - Functional workaround Tracking: https://github.com/github/github-mcp-server/issues/595

Issue #2: dbt SQL Execution Security

Server: dbt-mcp Impact: SQL execution tools can MODIFY data Mitigation: DISABLED by default via DISABLE_SQL=true, requires PAT Confidence: MEDIUM (0.65-0.70) when enabled - requires validation Recommendation: Keep disabled unless explicitly needed

Issue #3: Snowflake Cortex Tools Optional

Server: snowflake-mcp Impact: Cortex AI tools require explicit setup in Snowflake Mitigation: Document when Cortex features are available Confidence: MEDIUM (0.70-0.75) - Requires Cortex service configuration

Issue #4: aws-api READ_OPERATIONS_ONLY

Server: aws-api Impact: Cannot modify infrastructure via MCP (read-only mode) Mitigation: Design pattern - specialists recommend, human implements Confidence: Appropriate restriction - aws-expert provides guidance, not execution

Future MCP Integration Opportunities

High-Priority Additions

1. tableau-mcp (if available)

   • BI dashboard metadata
   • Data source analysis
   • Performance optimization
   • Primary User: tableau-expert, bi-developer-role

2. prefect-mcp (custom integration)

   • Flow orchestration
   • Deployment management
   • Run monitoring
   • Primary User: prefect-expert, data-engineer-role

3. orchestra-mcp (custom integration)

   • Pipeline orchestration
   • Workflow dependencies
   • Resource management
   • Primary User: orchestra-expert, data-engineer-role

Medium-Priority Additions

4. dbt Cloud Admin API (enhanced dbt-mcp)

   • Environment management
   • User/group administration
   • Account-level configuration
   • Primary User: dba-role, data-architect-role

5. Confluence MCP (knowledge management)

   • Team documentation
   • Meeting notes
   • Runbook management
   • Primary User: documentation-expert, business-analyst-role

Maintenance & Updates

When to Update This Guide

• New MCP server added to .mcp.json
• Specialist agent MCP tools updated
• New integration patterns validated in production
• Known issues resolved or new issues discovered
• Confidence scores change based on production validation

Update Protocol

1. Research: Comprehensive tool inventory and documentation
2. Document: Create detailed capability reference
3. Integrate: Update specialist agent files with MCP tools section
4. Test: Validate patterns with actual MCP calls
5. Update Guide: Add to this consolidated reference
6. Communicate: Update team on new capabilities

Verification Checklist

• All MCP servers documented with complete tool inventory
• Specialist agents updated with their MCP access
• Confidence scores assigned based on validation
• Known issues documented with workarounds
• Integration patterns validated in production
• Authentication and security documented
• Role agent delegation patterns defined

Quick Reference

MCP Server Status Check

claude mcp list

Test MCP Tool Availability

# Test dbt-mcp
mcp__dbt-mcp__list_metrics

# Test snowflake-mcp
mcp__snowflake-mcp__list_objects object_type="table" database_name="ANALYTICS_DW"

# Test aws-api
mcp__aws-api__call_aws cli_command="aws sts get-caller-identity"

# Test github
mcp__github__search_repositories query="org:your-org" perPage=5

Documentation Locations

• MCP Research: knowledge/mcp-servers/
• Specialist Agents: .claude/agents/specialists/
• Role Agents: .claude/agents/roles/
• Integration Patterns: .claude/rules/ and .claude/skills/reference-knowledge/
• This Guide: .claude/skills/reference-knowledge/agent-mcp-integration-guide/SKILL.md

Guide created: 2025-10-08 Author: Claude Code - MCP Integration Research Purpose: Complete MCP integration reference for DA Agent Hub agents

Converted and distributed by TomeVault — claim your Tome and manage your conversions.