Skip to content

03-agent-configuration-tool-governance-and-memory.md

Latest commit

 

History

History
134 lines (95 loc) · 5.03 KB

03-agent-configuration-tool-governance-and-memory.md

File metadata and controls

134 lines (95 loc) · 5.03 KB
layout default
title Chapter 3: Agent Configuration, Tool Governance, and Memory
nav_order 3
parent MCP Use Tutorial

Chapter 3: Agent Configuration, Tool Governance, and Memory

Welcome to Chapter 3: Agent Configuration, Tool Governance, and Memory. In this part of MCP Use Tutorial: Full-Stack MCP Development Across Agents, Clients, Servers, and Inspector, you will build an intuitive mental model first, then move into concrete implementation details and practical production tradeoffs.

Agent reliability depends on explicit control of tools, memory, and step budgets.

Learning Goals

  • configure MCPAgent with practical limits (maxSteps, memory)
  • apply disallowedTools to reduce unsafe or irrelevant tool use
  • use server-manager patterns for multi-server environments
  • align LLM selection with tool-calling support expectations

Governance Pattern

  1. start with minimal tool surface
  2. explicitly block dangerous categories unless needed
  3. set conservative step limits first
  4. monitor behavior before widening capability scope

Source References

Summary

You now have agent-level guardrails for safer, more predictable tool execution.

Next: Chapter 4: TypeScript Server Framework and UI Widgets

Source Code Walkthrough

libraries/python/mcp_use/logging.py

The provides class in libraries/python/mcp_use/logging.py handles a key part of this chapter's functionality:

Logger module for mcp_use.

This module provides a centralized logging configuration for the mcp_use library,
with customizable log levels and formatters.
"""

import logging
import os
import sys

from langchain_core.globals import set_debug as langchain_set_debug

# Global debug flag - can be set programmatically or from environment
MCP_USE_DEBUG = 1


class Logger:
    """Centralized logger for mcp_use.

    This class provides logging functionality with configurable levels,
    formatters, and handlers.
    """

    # Default log format
    DEFAULT_FORMAT = "%(asctime)s - %(name)s - %(levelname)s - %(message)s"

    # Module-specific loggers
    _loggers = {}

    @classmethod
    def get_logger(cls, name: str = "mcp_use") -> logging.Logger:
        """Get a logger instance for the specified name.

This class is important because it defines how MCP Use Tutorial: Full-Stack MCP Development Across Agents, Clients, Servers, and Inspector implements the patterns covered in this chapter.

libraries/python/examples/example_middleware.py

The TimingMiddleware class in libraries/python/examples/example_middleware.py handles a key part of this chapter's functionality:

    # Create custom middleware
    class TimingMiddleware(Middleware):
        async def on_request(self, context: MiddlewareContext[Any], call_next: NextFunctionT) -> Any:
            start = time.time()
            try:
                print("--------------------------------")
                print(f"{context.method} started")
                print("--------------------------------")
                print(f"{context.params}, {context.metadata}, {context.timestamp}, {context.connection_id}")
                print("--------------------------------")
                result = await call_next(context)
                return result
            finally:
                duration = time.time() - start
                print("--------------------------------")
                print(f"{context.method} took {int(1000 * duration)}ms")
                print("--------------------------------")

    # Middleware that demonstrates mutating params and adding headers-like metadata
    class MutationMiddleware(Middleware):
        async def on_call_tool(self, context: MiddlewareContext[Any], call_next: NextFunctionT) -> Any:
            # Defensive mutation of params: ensure `arguments` exists before writing
            try:
                print("[MutationMiddleware] context.params=", context.params)
                args = getattr(context.params, "arguments", None)
                if args is None:
                    args = {}

                # Inject a URL argument (example) and a trace id
                args["url"] = "https://github.com"
                meta = args.setdefault("meta", {})

This class is important because it defines how MCP Use Tutorial: Full-Stack MCP Development Across Agents, Clients, Servers, and Inspector implements the patterns covered in this chapter.

How These Components Connect

flowchart TD
    A[provides]
    B[TimingMiddleware]
    A --> B
Loading