lazy-mcp

Overview

Lazy MCP is a sophisticated proxy server that implements hierarchical tool organization with lazy loading capabilities. It intercepts MCP tool requests and only activates servers when specific tools are actually needed, significantly reducing context pollution and token consumption. The server exposes two meta tools: 'get_tools_in_category' for browsing tool hierarchies and 'execute_tool' for on-demand tool execution. This approach can save substantial tokens, as demonstrated in their example where it reduced context usage by 17% (34,000 tokens).

Try asking AI

After installing, here are 5 things you can ask your AI assistant:

• get_tools_in_categoryget_tools_in_category(path)

  Navigate the tool hierarchy to explore available tools and categories

  Try:What tools are available in the coding_tools category?Try:Show me all available categoriesTry:What tools are under web_tools?
• execute_toolexecute_tool(tool_path, arguments)

  Execute tools by their hierarchical path with lazy loading support

  Try:Find symbol in coding toolsTry:Create a GitHub issueTry:Run the serena tool to find symbols

Comparable tools

Installation

git clone https://github.com/voicetreelab/lazy-mcp.git
cd lazy-mcp

make build

./build/structure_generator --config config.json --output testdata/mcp_hierarchy

claude mcp add --transport stdio mcp-proxy build/mcp-proxy -- --config config.json

FAQ

How does lazy loading save tokens?

Lazy MCP only activates MCP servers when specific tools are actually requested, rather than loading all tools at startup. This keeps unused tool instructions out of the context window, significantly reducing token consumption.

Can I control which tools require permission?

Yes, Lazy MCP supports multiple permission control methods: Claude Code native hooks, token-based hooks, and an OpenCode plugin. These allow you to define sensitive tools that require confirmation before execution.

Compare lazy-mcp with