mcp-client

MCP client connects to MCP servers using HTTP or stdio transports to dynamically discover and execute tools on-demand. This approach lists available tools from the server, caches their schemas for reference, and invokes them as needed—reducing token usage by avoiding upfront loading of all tool definitions into your AI skill's context.

Yes, MCP client supports both HTTP transports for remote MCP servers and stdio transports for local servers. This flexibility allows you to work with MCP servers running anywhere in your architecture while maintaining a single unified client interface for tool discovery and invocation.

MCP client implements JSON-RPC communication by initializing MCP sessions, establishing transports, and handling message exchange automatically. After copying the mcp-client.py script into your skill and configuring the server URL, the client manages all protocol details for listing and calling tools.

MCP client generates and caches tool schemas to the references/ directory, enabling progressive disclosure of tool definitions. Schema caching keeps your workflow modular and reduces the number of tool definitions loaded into context at any given time, supporting on-demand tool discovery.

No, MCP client eliminates manual tool loading by dynamically discovering tools from MCP servers. Tools are listed and invoked on-demand through the client, with schemas cached for reference—this keeps tool definitions lightweight and out of your skill's main context.

Yes, MCP client provides a unified interface for tool invocation and result processing regardless of server count or type. The single client abstracts transport and JSON-RPC details, letting you call tools and handle responses uniformly across HTTP and stdio MCP servers.