mcpcute

MCP aggregator - aggregate multiple MCPs behind a single interface to reduce context pollution for AI agents.

Instead of exposing 20+ MCP tools directly to your AI agent, mcpcute provides a two-level hierarchy with just 7 tools:

MCP-Level Operations

1. list_mcps - List all available MCP servers with their connection status

2. search_mcps - Search for MCP servers by name

3. get_mcp_details - Get detailed info about an MCP including its tools

Tool-Level Operations

4. list_tools - List all tools for a specific MCP

5. search_tools - Search for tools (optionally scoped to a specific MCP)

6. get_tool_details - Get detailed schema and description for a tool

7. execute_tool - Execute a tool from a chosen MCP (requires both MCP and tool names plus arguments)

Installation

Configuration

Create a mcpcute.config.json file in your working directory (or set MCPCUTE_CONFIG env var to point to your config file):

Usage

With Claude Desktop

Add to your claude_desktop_config.json:

With Claude Code

Add to your Claude Code MCP settings:

Standalone

How it works

1. mcpcute starts instantly - no upfront connections to any MCP servers

2. Use list_mcps or search_mcps to discover available MCPs (no connections needed)

3. Use get_mcp_details or list_tools to explore an MCP's capabilities (connects on-demand)

4. Use search_tools to find tools across all MCPs or scoped to one

5. Use get_tool_details to get the full schema for a tool

6. Use execute_tool (with mcp_name, tool_name, and arguments) to run the tool

This reduces the initial context from potentially hundreds of tool schemas to just 7 simple tools, and startup is instant regardless of how many MCPs you configure.

Cache

mcpcute persists the discovered tool list for each MCP to disk so future runs can answer tool discovery requests without reconnecting to every server. The cache lives in:

• macOS/Linux: ${XDG_CACHE_HOME:-~/.cache}/mcpcute

• Windows: %LOCALAPPDATA%/mcpcute/cache

Override the location with MCPCUTE_CACHE_DIR. Cached entries are automatically invalidated whenever the command, arguments, or environment for a server change in mcpcute.config.json.

Workflow Examples

Discovering filesystem tools

Exploring all available MCPs

Why mcpcute?

• Instant startup: Lazy loading means no waiting for 20+ MCP servers to connect

• Two-level hierarchy: Clear separation between MCP discovery and tool discovery

• Reduced context pollution: Instead of loading 50+ tool schemas into your AI's context, load just 7

• Dynamic tool discovery: AI agents can search and discover tools as needed

• Scoped exploration: Explore one MCP at a time instead of being overwhelmed

• Unified interface: One consistent API for all your MCP tools

• Easy configuration: Simple JSON config to aggregate multiple MCP servers

License

MIT