16. Claude Code Integration

Chapter 16 of 22 · 20 min

KEY INSIGHT

MCP servers extend Claude Code's capabilities by exposing local tools as native featuresΓÇöenabling AI to interact with your systems using a standardized protocol. Claude Code communicates with MCP servers throughstdio transport by default. Installation requires registering the server in Claude Code's configuration: ```json { "mcpServers": { "filesystem": { "command": "uv", "args": ["run", "mcp-server-filesystem", "--path", "/projects"] }, "database": { "command": "docker", "args": ["run", "--rm", "-e", "DB_URL=postgres://localhost/mydb", "mcp-database-server"] } } } ``` After configuration, Claude Code discovers available tools automatically. Use `/mcp` to see which servers are running and what tools they provide. Create Claude Code-optimized tools by understanding how the model interprets them. Tool descriptions become part of the promptΓÇöwrite them as instructions, not API documentation: ```python @mcp.tool(description="""Read the contents of a file. Args: - path: Absolute path to the file (required) Returns the complete file contents as text. Fails if the file exceeds 1MB or is a binary format. Example use: - Reading source code before editing - Examining configuration files - Reviewing logs to debug issues""") def read_file(path: str) -> str: ... ``` Schema definitions directly influence how Claude Code requests tool execution. Use descriptive field names and provide constraints: ```python @mcp.tool(description="Search for files matching criteria") async def search_files( pattern: Annotated[str, Field(description="Glob pattern, e.g. '*.py'")], directory: Annotated[str, Field( description="Root directory for search", default="/projects" )], case_sensitive: Annotated[bool, Field( description="Match case exactly", default=False )], ) -> list[dict]: ... ``` For complex operations, consider tools that return structured data rather than raw text. Claude Code processes structured responses more reliably: ```python @mcp.tool(description="Query the database with SQL") async def query_database( sql: Annotated[str, Field(description="SELECT statement (read-only)")], ) -> dict: """Execute query and return structured results.""" if not sql.strip().upper().startswith("SELECT"): raise ValueError("Only SELECT statements allowed") results = await execute_query(sql) return { "columns": list(results[0].keys()) if results else [], "rows": [dict(row) for row in results], "count": len(results), } ``` Resources provide read-only data access that Claude Code can query without tool invocation. Register resource providers for static or frequently-accessed data: ```python @mcp.resource("config://app/settings") def get_app_settings() -> str: return json.dumps(AppSettings.current().to_dict()) @mcp.resource("file://{path}") def get_file_resource(path: str) -> str: return Path(path).read_text() ```

EXERCISE

Create an MCP server with three tools optimized for Claude Code interaction: one for file operations, one for running shell commands, one for reading documentation. Test each tool through Claude Code using the MCP configuration.