mai-yyy/multi-llm-mcp

This is a plugin (called an MCP server) for Claude Code that adds two main capabilities. The README is written in Chinese. First, it lets Claude Code hand off coding tasks to OpenAI's Codex CLI, which is a separate AI coding agent. Second, it lets Claude Code send the same question to multiple AI models at once, such as GPT, DeepSeek, Kimi, and Qwen, so you can compare their answers side by side. The Codex integration works by running Codex as a background job. Because some coding tasks take a while to complete and MCP tool calls have a timeout limit, the plugin returns a job ID immediately and then provides a separate tool you call later to retrieve the result when it is ready. Codex can be run in three permission levels: read-only for browsing and analyzing code, workspace-write for making file changes, and a higher-access mode the documentation advises against using casually. The multi-model feature works similarly. You can ask a single model a question in a conversational session that preserves context across multiple turns, or you can broadcast the same question to several models simultaneously and get all their responses back. Long-running multi-model requests also use the job ID pattern to avoid timeout problems. A health check tool lets you verify which models are configured and available without exposing any API keys. Each AI provider needs its own API key set as an environment variable. The plugin connects to DeepSeek, Kimi (Moonshot), Qwen (via DashScope), OpenAI, and optionally Claude via Anthropic's API. It is built on the FastMCP framework, has no other dependencies beyond the OpenAI Python SDK, and is installed by registering the Python script as an MCP server in Claude Code's configuration. The license is MIT.