Plugins & MCP Overview

The plugin system extends agent capabilities through two mechanisms: MCP (Model Context Protocol) server integration and Python plugin files.

Components

Module	Purpose
MCP Adapter	Wrapping MCP tools as native ToolCapability
Plugin Loader	Python plugin files with contribute() functions

MCP Integration

MCP (Model Context Protocol) allows agents to connect to external tool servers that provide additional capabilities. langgraph-kit wraps MCP tools into the native ToolCapability model so they integrate seamlessly with the tool registry, risk filtering, and prompt assembly.

MCP Server (external process)
    │
    ▼
MCPClientManager (manages connections)
    │
    ▼
MCPToolAdapter (wraps as ToolCapability)
    │
    ▼
ToolRegistry (standard registration)
    │
    ▼
Agent can use MCP tools like any other tool

Python Plugins

For extensions that don't need a separate server process, Python plugin files provide a simpler mechanism:

# plugins/my_plugin.py
def contribute():
    """Called during agent startup."""
    return PluginContribution(
        tools=[my_custom_tool],
        sections=[my_prompt_section],
    )

Configuration

MCP servers are configured via the mcp_servers field in AgentConfig as a JSON string:

AgentConfig(
    mcp_servers='[{"name": "filesystem", "command": "npx", "args": ["@modelcontextprotocol/server-filesystem", "/path"]}]'
)

Python plugins are loaded from the plugins_dir directory:

AgentConfig(plugins_dir="/path/to/plugins/")