mcp-client_skill

Overview

This skill is a universal MCP client that connects to any MCP server with progressive disclosure. It loads tool schemas on-demand to avoid bloating the conversation context. Use it to list available MCP servers, inspect tool schemas, and execute remote MCP tool calls safely from a single Python script.

How this skill works

The skill reads configuration from several locations (MCP_CONFIG_PATH, a local reference config, .mcp.json, or ~/.claude.json) to discover servers and transport types. It supports HTTP (FastMCP with Bearer auth), stdio command transports, SSE, and streamable HTTP endpoints, and fetches tool schemas only when you request them. If a required credential (for example, a Zapier API key) is missing, the script will surface that and you should provide it in the config.

When to use it

• Connect to an external MCP server such as Zapier, Sequential Thinking, GitHub, or a filesystem adapter.
• Inspect available tools and their full parameter schemas without loading everything into your chat context.
• Execute a specific MCP tool call with validated JSON arguments.
• Confirm which MCP servers are configured in your environment before calling tools.
• Debug connection or configuration issues with clear error types.

Best practices

• Keep credentials in the recommended config file (references/mcp-config.json) or set MCP_CONFIG_PATH to a secure path.
• List servers first to confirm configuration before calling tools (python ... mcp_client.py servers).
• List tool schemas on-demand to avoid context bloat (python ... mcp_client.py tools <server>).
• Validate JSON arguments against the displayed schema before calling tools to reduce runtime errors.
• Install dependencies in a virtualenv: pip install mcp fastmcp.

Example use cases

• Connect to Zapier, list available Zapier actions, then call a specific Zapier action with JSON params.
• Run sequential-thinking via stdio transport to break a problem into numbered thoughts.
• Query a GitHub-backed MCP server to list repository automation tools and trigger a workflow.
• Use an SSE MCP endpoint for streaming tool responses while keeping tool schemas out of the chat context.
• Use the servers command to audit what MCP backends are available in your environment.