joeseesun/qiaomu-llm-mcp

Free

把多模型 Provider、本地密钥统一成一个 MCP 网关，为多供应商 LLM 提供统一路由与密钥管理，AI 客户端一处接入即可调用多家模型。

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About joeseesun/qiaomu-llm-mcp

Qiaomu-LLM-MCP is a local MCP (Model Context Protocol) gateway that unifies multiple LLM providers, API keys, routing rules, and multi-model discussions into a single tool surface. It allows AI clients like Codex, Claude Desktop, or Claude Code to access a variety of models through one MCP server. The tool provides a set of tools: listing providers, viewing provider configs, listing models, routing tasks by type, single-model chat, multi-model comparison, model pipeline, and HeavySkill-style multi-round discussions with a moderator. API keys are securely stored in the macOS Keychain or environment variables, and the server never exposes plaintext keys. The provider registry is saved locally at ~/.config/qiaomu-llm/registry.json. It supports Z.ai, DeepSeek, OpenAI-compatible, Anthropic-compatible, and Claude Code CLI. This project is local-first and open source.

Key Features

Unifies multiple LLM providers (Z.ai, DeepSeek, OpenAI-compatible, Anthropic-compatible, Claude Code CLI) under a single MCP surface

Secure API key storage via macOS Keychain or environment variables; server never exposes plaintext keys

Local provider registry at ~/.config/qiaomu-llm/registry.json, easy to manage

Tools include: qllm_list_providers, qllm_show_provider, qllm_list_models, qllm_route_task, qllm_chat, qllm_compare, qllm_pipeline, qllm_heavy_discuss, qllm_heavy_render, qllm_claude_code_*

Intelligent task routing: selects provider/model based on task type without consuming tokens

Multi-model comparison and pipeline (chain multiple models) for complex workflows

HeavySkill-style multi-round discussion with a designated moderator model, outputting Markdown/HTML reports

Works with MCP-compatible clients like Codex, Claude Desktop, Claude Code, Cursor

Pros & Cons

Pros

Centralizes multiple LLM providers into one MCP gateway, simplifying client integration
Secure key management: uses macOS Keychain, never exposes API keys in plaintext
Supports advanced workflows: single chat, comparison, pipeline, and multi-model discussions
Task routing chooses appropriate model without extra token cost
Local-first and open source, giving users full control over their data and configuration
Compatible with popular MCP clients like Codex, Claude Desktop, and Claude Code

Cons

Currently only tested on macOS and Linux; no Windows support mentioned
Requires Python 3.10+ and manual setup of provider registry and keychain entries
Not a turnkey SaaS solution; users must install and configure locally
HeavySkill discussion feature depends on availability of a strong moderator model
Learning curve for users unfamiliar with MCP or command-line tools

Best For

Developers using multiple LLM providers (e.g., Z.ai, DeepSeek, OpenAI, Anthropic) who want a single integration pointTeams that want to avoid hardcoding API keys in scripts, prompts, or configuration filesRunning multi-model comparisons or pipelines from within MCP clients (Codex, Claude, etc.)Conducting structured multi-perspective discussions where different models take different roles (e.g., initial draft, review, final expression)Centralizing provider and key management for AI-assisted development workflows

FAQ

Which LLM providers are supported?

The tool supports Z.ai (Coding Plan), DeepSeek, OpenAI-compatible APIs, Anthropic-compatible APIs, and Claude Code CLI. The provider registry is fully customizable.

How are API keys stored and secured?

API keys are stored in the macOS Keychain by default, referenced via secret_ref like 'keychain:qiaomu-llm/zai-glm'. Environment variables are also supported but Keychain is recommended for production use.

What is HeavySkill discussion?

HeavySkill discussion is a multi-model, multi-perspective conversation where different models provide independent opinions on a topic. A moderator (either a model or Codex) synthesizes the discussion, and the result can be rendered as a Markdown or HTML report.

Which MCP clients can use this gateway?

Any MCP-compatible client can connect, including Codex, Claude Desktop, Claude Code CLI, and Cursor. Configuration typically involves adding the server to the client's MCP settings.

Is the tool free and open source?

Yes, qiaomu-llm-mcp is open source under the license in the repository. There are no usage fees, though users pay for their own API keys from providers.