Model Context Protocol (MCP)

The Model Context Protocol (MCP) forces a separation between the definition of a tool (the server) and the consumption of a tool (the client/agent). Instead of hardcoding API integrations inside your agent's codebase, you build an MCP Server that exposes resources (data) and tools (functions). Any MCP-compliant agent can then discover and use these tools without custom glue code.

When to Use

• Standardization: When building an ecosystem of tools that many different agents need to use.
• Security: To expose internal data safely without giving the LLM direct database access.
• Modularity: To keep your agent logic clean and focused on reasoning, while the MCP server handles the "dirty work" of API connections.
• Portability: Tools built with MCP can be used by Claude Desktop, IDEs, and custom agents alike.

Use Cases

• Database Access: An MCP server that exposes safe SQL queries as tools.
• File System: An MCP server that allows an agent to read/write files in a sandboxed directory.
• API Wrapper: An MCP server that wraps the GitHub API, exposing actions like create_issue or list_prs.

Implementation Pattern

# MCP Server Implementation (Conceptual)
from fastmcp import FastMCP, tool

# Create a server
mcp = FastMCP("MyTools")

# Expose a tool
@mcp.tool()
def calculate_vat(amount: float, country: str) -> float:
    """Calculates VAT for a given country."""
    rate = get_rate(country)
    return amount * rate

# The Agent (Client) simply connects to this server
# and automatically "sees" the calculate_vat tool available for use.
# client.connect(mcp_server)
# response = client.chat("How much VAT for 100 EUR in Germany?")

Examples

Input: "Connect my agent to a database MCP server."

// .claude/settings.json
{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": { "DATABASE_URL": "postgresql://localhost/mydb" }
    }
  }
}

Agent can now call mcp__postgres__query to run SQL and mcp__postgres__schema to inspect tables.

---

Input: "Build a multi-MCP workflow: search the web, then save results to Notion."

Pattern: Phase 1 calls mcp__brave__search, collects results. Phase 2 calls mcp__notion__createPage with formatted output. Each phase validates data before proceeding.

Troubleshooting

Problem	Cause	Fix
MCP server not found	Binary not installed	Run npx -y @modelcontextprotocol/server-name to install
Tool calls return auth errors	Missing API key in env	Add key to env block in mcpServers config
Agent can't find MCP tools	Server not registered	Check .claude/settings.json; restart Claude Code after changes
Tool call times out	MCP server unresponsive	Test server independently: npx @modelcontextprotocol/inspector
Data format mismatch between MCPs	Different schemas	Add a transformation step between phases to normalize formats