agent-society-mcp-client-wrapper 0.0.1

A Python Library that wrapps around the low-level mcp client lib by Anthropic

MCP Wrapper for AgentSociety

A library that wraps low-level MCP client functionality, exposing it for easy usage as tools within a LangChain workflow. This enables seamless integration of remote MCP tools for language model workflows and agent applications.

Structure

Library Code: mcpwrap/

• client.py — Basic utility to invoke a LangChain model with a set of tools.
• server_stub.py — Async abstraction (ServerSession) for connecting to and interfacing with an MCP server, exposing MCP tools for LangChain.
• multi_mcp_model.py — Orchestrates multiple MCP server sessions and integrates their toolsets for use by one BaseChatModel. Handles async invocation and batching of tool requests.
• llm_integration.py — Converts MCP tool schemas (JSON Schema) into dynamic, structured LangChain tools using Pydantic models.
• (You may need to install mcp and LangChain dependencies.)

Example App: sample/

• Provides a minimal, integration-tested example of how to use this library in practice. (See directory for details.)

Usage Overview

1. Connect MCP servers as tool providers

from mcpwrap.server_stub import ServerSession

session = ServerSession(name="example", url="http://localhost:8080")
await session.initialize()  # Async context

2. Compose Multi-server Tools for LangChain

from langchain_core.language_models.chat_models import BaseChatModel
from mcpwrap.multi_mcp_model import MultiMcpModel

base_model = ... # Any supported BaseChatModel
mcp_model = MultiMcpModel(
    base_model=base_model,
    mcp_servers=[session]  # Add as many as you like
)
await mcp_model.initialize()

3. Use in a Chat Workflow

results = await mcp_model.ainvoke(messages)  # messages: List[BaseMessage]

Features

• Wraps multiple MCP servers: Maps their tools to a common namespace for use as LangChain tools.
• Schema Conversion: Converts arbitrary MCP tool JSON schemas into Pydantic models for argument validation.
• Async, batched operations: Tools are called concurrently when invoked by an LLM agent.
• Error and retry handling: Optionally retries failed server/tool executions.

Requirements

• pip install -r requirements.txt

Notes

• Ensure MCP servers are reachable at their configured URLs.
• Schema support currently assumes JSON Schema Draft 7 compatibility for tool arguments.

Contributing

PRs, issues, and questions are welcome!