MCP Servers Essential Reference

Skill by ara.so — MCP Skills collection.

This skill provides expertise in selecting, installing, and configuring MCP (Model Context Protocol) servers for AI coding agents like Claude Code, Cursor, Codex, and Gemini CLI. It covers 50 essential servers across 9 categories with installation commands, security best practices, and troubleshooting.

What MCP Servers Are

The Model Context Protocol (MCP) is a standardized interface that allows AI agents to connect to external tools and services. Instead of each AI model implementing custom integrations, MCP provides a "USB-C port for AI" — one protocol that both agents and tools speak.

Key concepts:

• Server: A service that exposes tools/resources via MCP (e.g., GitHub API, database access)
• Transport: How the agent connects (stdio for local, HTTP/SSE for hosted)
• Tool: A function the agent can call (e.g., search_github, run_sql_query)
• Resource: Read-only data the agent can access (e.g., file contents, API docs)

The Three Critical Rules

Before installing any MCP server:

1. Don't install more than 5-7 servers — Tool bloat degrades agent performance and reasoning quality
2. Treat every server like untrusted code — Pin versions, scope tokens read-only, review source
3. Never enable writes to production — Point at replicas, snapshots, or staging environments

Installation Methods by Agent

Claude Code (Desktop)

# Local server (stdio transport)
claude mcp add github -- npx -y @modelcontextprotocol/server-github

# Hosted server (HTTP transport)
claude mcp add linear --transport http

# List installed servers
claude mcp list

# Remove a server
claude mcp remove github

Config file: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)

Codex

# Add server
codex mcp add postgres -- npx -y @modelcontextprotocol/server-postgres

# Edit config directly
vim ~/.codex/config.toml

Gemini CLI

Edit ~/.gemini/settings.json:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

Manual Configuration Format

All agents use similar JSON structure for mcpServers:

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "@scope/package-name"],
      "env": {
        "API_KEY": "${ENV_VAR_NAME}"
      }
    }
  }
}

For hosted servers:

{
  "mcpServers": {
    "linear": {
      "url": "https://mcp.linear.app",
      "transport": "http"
    }
  }
}

Category-Based Server Selection

Core Development (Pick 3-4)

GitHub (Official, hosted)

claude mcp add github -- npx -y @modelcontextprotocol/server-github

Environment: GITHUB_PERSONAL_ACCESS_TOKEN

Capabilities: Search repos/issues/PRs, create issues, manage PRs, push commits

Filesystem (Official, local)

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/directory

⚠️ Security: Only expose specific directories, never / or ~

Brave Search (Official, local)

claude mcp add brave-search -- npx -y @modelcontextprotocol/server-brave-search

Environment: BRAVE_API_KEY

Use case: Web search without leaving the agent context

Playwright (Official, local)

claude mcp add playwright -- npx -y @executeautomation/playwright-mcp-server

Capabilities: Browser automation, screenshot capture, web scraping

Databases & Backend (Pick 1)

PostgreSQL (Official, local)

claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres postgresql://user:pass@localhost/dbname

⚠️ Critical: Use read-only user or point at replica

Supabase (Official, hosted)

claude mcp add supabase --transport http

Requires: Supabase project URL and service role key

SQLite (Community, local)

claude mcp add sqlite -- npx -y @benborla/mcp-server-sqlite /path/to/database.db

Ops & Infrastructure

AWS (Official, local)

claude mcp add aws -- npx -y @modelcontextprotocol/server-aws-kb-retrieval-server

Environment: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION

Cloudflare (Official, hosted)

claude mcp add cloudflare --transport http

⚠️ Dangerous: Can modify DNS, workers, R2 buckets

Kubernetes (Community, local)

claude mcp add kubernetes -- npx -y @strowk/mcp-k8s-go

Environment: KUBECONFIG path ⚠️ Production risk: Can delete pods, services, deployments

Productivity & Team Tools

Slack (Community, local)

claude mcp add slack -- npx -y @modelcontextprotocol/server-slack

Environment: SLACK_BOT_TOKEN, SLACK_TEAM_ID

Linear (Official, hosted)

claude mcp add linear --transport http

OAuth flow via browser

Notion (Official, hosted)

claude mcp add notion --transport http

Google Drive (Official, local - archived)

claude mcp add gdrive -- npx -y @modelcontextprotocol/server-gdrive

Requires OAuth credentials JSON

Fintech & Payments (⚠️ All flagged as money-moving)

Stripe (Official, hosted/local)

claude mcp add stripe -- npx -y @stripe/mcp

Environment: STRIPE_SECRET_KEY (use test keys first)

Plaid (Official, hosted) Documentation: Check Plaid developer portal for latest MCP endpoint

⚠️ All fintech servers: Start in sandbox/test mode, never use production keys until you've verified agent behavior

Web3 & Blockchain (⚠️ Irreversible transactions)

Base MCP (Official, hosted)

claude mcp add base --transport http

Capabilities: Deploy contracts, send transactions, query onchain data

⚠️ Critical: Transactions are irreversible. Use testnet first.

The Graph (Community, local - read-only)

claude mcp add thegraph -- npx -y @rainprotocol/mcp-subgraph

Safe for querying blockchain data via GraphQL

Dune (Community, local - read-only) Environment: DUNE_API_KEY Use case: SQL queries against blockchain data

Trading & Markets

Alpaca (Official, local)

claude mcp add alpaca -- npx -y @alpacahq/mcp-server

Environment: APCA_API_KEY_ID, APCA_API_SECRET_KEY, APCA_API_BASE_URL

⚠️ Paper trade first: Set APCA_API_BASE_URL=https://paper-api.alpaca.markets

Polygon.io (Official, local - read-only)

claude mcp add polygon -- npx -y @modelcontextprotocol/server-polygon

Environment: POLYGON_API_KEY

CoinGecko (Official, hosted - read-only) Safe for market data queries

Creator Tools & Media

ElevenLabs (Official, local)

claude mcp add elevenlabs -- npx -y @modelcontextprotocol/server-elevenlabs

Environment: ELEVENLABS_API_KEY

Figma (Official, hosted/local)

claude mcp add figma -- npx -y @modelcontextprotocol/server-figma

Environment: FIGMA_PERSONAL_ACCESS_TOKEN

YouTube (Community, local) Environment: YOUTUBE_API_KEY Capabilities: Search videos, retrieve metadata, manage playlists

Reasoning & Memory

Sequential Thinking (Official, local)

claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

Forces step-by-step reasoning for complex problems

Memory (Official, local)

claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

Persistent knowledge graph across sessions

Real-World Configuration Examples

Minimal Dev Setup (5 servers)

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects"]
    },
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://readonly:${DB_PASSWORD}@localhost/myapp_dev"]
    },
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "${BRAVE_API_KEY}"
      }
    },
    "sequential-thinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    }
  }
}

Creator Pipeline (4 servers)

{
  "mcpServers": {
    "elevenlabs": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-elevenlabs"],
      "env": {
        "ELEVENLABS_API_KEY": "${ELEVENLABS_API_KEY}"
      }
    },
    "figma": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-figma"],
      "env": {
        "FIGMA_PERSONAL_ACCESS_TOKEN": "${FIGMA_TOKEN}"
      }
    },
    "youtube": {
      "command": "npx",
      "args": ["-y", "@benborla/mcp-server-youtube"],
      "env": {
        "YOUTUBE_API_KEY": "${YOUTUBE_API_KEY}"
      }
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/content"]
    }
  }
}

Team Productivity (5 servers)

{
  "mcpServers": {
    "linear": {
      "url": "https://mcp.linear.app",
      "transport": "http"
    },
    "notion": {
      "url": "https://mcp.notion.so",
      "transport": "http"
    },
    "slack": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-slack"],
      "env": {
        "SLACK_BOT_TOKEN": "${SLACK_BOT_TOKEN}",
        "SLACK_TEAM_ID": "${SLACK_TEAM_ID}"
      }
    },
    "gdrive": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-gdrive"],
      "env": {
        "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/credentials.json"
      }
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

Security Best Practices

Token Scoping (GitHub Example)

Create read-only token for initial testing:

• Permissions: repo:status, public_repo (read)
• No write permissions until you've audited agent behavior

Upgrade to write access only when needed:

• Add: repo (full control)
• Monitor first 10-20 agent actions manually

Database Access Patterns

Create read-only user:

-- PostgreSQL
CREATE USER mcp_readonly WITH PASSWORD 'secure_password';
GRANT CONNECT ON DATABASE myapp TO mcp_readonly;
GRANT USAGE ON SCHEMA public TO mcp_readonly;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO mcp_readonly;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO mcp_readonly;

Connection string:

postgresql://mcp_readonly:${MCP_DB_PASSWORD}@localhost:5432/myapp

File System Isolation

{
  "filesystem": {
    "command": "npx",
    "args": [
      "-y",
      "@modelcontextprotocol/server-filesystem",
      "/Users/you/safe-workspace",
      "--readonly"
    ]
  }
}

Never expose:

• / (system root)
• ~ (home directory)
• /etc, /usr, /bin
• .ssh, .aws, .config directories

Troubleshooting

Server Won't Start

Check logs:

# Claude Desktop logs (macOS)
tail -f ~/Library/Logs/Claude/mcp-server-*.log

# Check if process is running
ps aux | grep npx

Common issues:

1. Missing environment variable: Check .env or shell profile
2. Port conflict: Another server using same port (hosted servers)
3. npx version mismatch: Update Node.js to LTS version
4. Permission denied: File system server needs directory read/write access

Fix:

# Reload environment
source ~/.zshrc  # or ~/.bashrc

# Clear npx cache
npx clear-npx-cache

# Reinstall server
claude mcp remove server-name
claude mcp add server-name -- npx -y @scope/package

Agent Can't See Tools

Verify server is connected:

claude mcp list
# Should show "connected" status

Check agent prompt:

• Some agents require explicit mention: "use the github server to..."
• Tool discovery happens on session start — restart the agent

Test with simple command:

• "List all available tools"
• "What can you do with the github server?"

Performance Degradation

Symptom: Agent responses slow, frequent timeouts

Diagnosis:

# Count active servers
claude mcp list | grep connected | wc -l

# Check context window usage (if agent provides)
# Claude: look for "context: 45% used" in UI

Fix: Disable servers you're not actively using

claude mcp remove server1
claude mcp remove server2
# Keep only 3-5 essential servers

Authentication Errors

GitHub "Bad credentials":

# Verify token has required scopes
curl -H "Authorization: token ${GITHUB_TOKEN}" https://api.github.com/user

# Regenerate token at https://github.com/settings/tokens
# Update environment variable
export GITHUB_TOKEN="ghp_newtoken"

Database "Connection refused":

# Check database is running
pg_isready -h localhost -p 5432

# Test connection manually
psql postgresql://user:pass@localhost/dbname

Hosted server OAuth failures:

• Clear browser cache/cookies
• Try OAuth flow in incognito window
• Check redirect URL matches agent's expected callback

Version Pinning (Recommended)

Avoid breaking changes by pinning versions:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": [
        "@modelcontextprotocol/server-github@0.5.0"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

Update workflow:

# Test in isolated config first
cp claude_desktop_config.json claude_desktop_config.json.backup

# Update one server
# Edit config: @scope/package@0.5.0 → @scope/package@0.6.0

# Restart agent and test core workflows

# If broken, revert:
mv claude_desktop_config.json.backup claude_desktop_config.json

Common Patterns

Conditional Server Loading

Use environment variable to toggle expensive servers:

{
  "mcpServers": {
    "playwright": {
      "command": "sh",
      "args": [
        "-c",
        "[ \"$ENABLE_BROWSER\" = \"true\" ] && npx -y @executeautomation/playwright-mcp-server || echo 'disabled'"
      ]
    }
  }
}

Usage:

# Enable for this session
export ENABLE_BROWSER=true
claude

Multi-Environment Configs

Maintain separate configs for different workflows:

# ~/.config/claude/
├── config-dev.json       # GitHub, Postgres, Filesystem
├── config-trading.json   # Polygon, CoinGecko, Alpaca (paper)
├── config-creator.json   # ElevenLabs, Figma, YouTube
└── config-prod.json      # Minimal read-only servers

# Symlink active config
ln -sf ~/.config/claude/config-dev.json ~/Library/Application\ Support/Claude/claude_desktop_config.json

Monitoring Agent Actions

For high-risk servers (databases, cloud infra), log all operations:

PostgreSQL audit log:

-- Enable query logging
ALTER SYSTEM SET log_statement = 'all';
ALTER SYSTEM SET log_directory = 'pg_log';
SELECT pg_reload_conf();

-- Monitor in real-time
tail -f /var/lib/postgresql/data/pg_log/postgresql-*.log

GitHub webhook for audit trail: Create webhook to log all agent-initiated changes to separate audit system.

Quick Reference: 50 Servers by Risk Level

✅ Safe (Read-only, no credentials)

• Sequential Thinking, Memory, Brave Search (with API key)

⚠️ Low Risk (Read-only with auth)

• GitHub (read scope), Linear, Notion, Sentry, Polygon.io, CoinGecko, The Graph, Dune

🟡 Medium Risk (Can modify data)

• Filesystem (scoped), Slack, Gmail, Google Calendar, Figma, ElevenLabs, YouTube

🔴 High Risk (Infrastructure/money)

• Databases: Postgres, Supabase, Neon, SQLite, Redis (write access)
• Cloud: AWS, Cloudflare, Vercel, Kubernetes, Docker Hub
• Fintech: Stripe, PayPal, Plaid, QuickBooks
• Trading: Alpaca, CCXT
• Web3: Base, Solana Agent Kit, Thirdweb (testnet OK)

Rule: Start all 🔴 servers in read-only/sandbox/testnet mode. Graduate to write access only after 20+ successful manual-approved operations.

Resources

• Official MCP Docs: https://modelcontextprotocol.io
• Server Registry: https://github.com/modelcontextprotocol/servers
• Security Audit: Review 2026 Q1 security findings before installing community servers
• SDK Downloads: npm @modelcontextprotocol/sdk (~97M monthly downloads as of March 2026)

Remember: This is a menu, not a mandate. Pick 3-5 servers that match your actual workflow, secure them properly, and ship.