mcp-client

You can connect Claude to an MCP server by loading tool schemas on demand, which keeps prompts lean and avoids context window bloat. This approach supports listing servers, enumerating tools, and invoking actions across Zapier, GitHub, and local transports.

You can use stdio, SSE, streamable HTTP, and FastMCP transports to run MCP tool invocations. The client provides unified commands to connect, list tools with full schemas, and call tools across these different transport types.

Yes, you need the Python dependencies mcp and fastmcp installed to operate the MCP client. You also need a configured MCP setup such as MCP_CONFIG, a .mcp.json file, or an equivalent configuration to connect to external servers.

To list tools, run the client command targeting Zapier, which enumerates tools with full schemas. To call a tool, use the invoke command with the server name, tool name, and a JSON string of parameters to trigger the action.

Loading all MCP tool schemas simultaneously consumes significant context window space because each tool includes full parameter definitions. On-demand loading solves this by only fetching schemas when you need to invoke a specific tool.

Yes, you can run local MCP servers for sandboxed tasks using supported transports like stdio. The client connects to these local servers to discover available tools and trigger actions within your sandboxed environment.