Claude Concilium

What spyrae/claude-concilium MCP server does

The spyrae/claude-concilium MCP server coordinates multiple LLM consultations for Claude Code. It contains separate MCP servers for OpenAI through the Codex CLI, Gemini through gemini-cli, and Qwen through the Qwen CLI. These servers let an agent send a problem to one or more external model providers and compare their responses.

The repository is intended for situations where a second opinion may expose a bug, edge case, architectural concern, or alternative solution. Documented examples include code review, architecture decisions, stalled debugging work, performance optimization, and security review. The project also provides a Claude Code command skill that can initiate a multi-agent consultation.

How it works

A typical consultation starts with a concise problem description, documented as being under 500 characters. Claude Code can call the OpenAI and Gemini MCP tools in parallel, then compare the returned opinions. If responses disagree, the workflow can continue with follow-up questions before a final decision is made.

The spyrae/claude-concilium MCP server uses fallback chains when a provider reports a quota issue, authentication problem, unsupported model, or timeout. The documented path can move from OpenAI or Gemini to Qwen, then to DeepSeek. Failed CLI processes can be automatically terminated when they exceed the timeout. Errors are returned in structured form, including identifiers such as QUOTA_EXCEEDED, AUTH_EXPIRED, AUTH_REQUIRED, AUTH_NOT_CONFIGURED, and MODEL_NOT_SUPPORTED.

Setup and configuration

The project requires Node.js 20 or newer. Installation is performed separately in each server directory: servers/mcp-openai, servers/mcp-gemini, and servers/mcp-qwen. After installing dependencies, the repository provides a smoke test that checks all three servers without requiring the provider CLIs.

To connect the servers to Claude Code, copy config/mcp.json.example to .mcp.json and replace its example paths with absolute paths on the local machine. Servers may also be added individually. The optional skill is installed by copying skill/ai-concilium.md to ~/.claude/commands/ai-concilium.md, after which /ai-concilium can be used in Claude Code.

Provider setup is separate from MCP installation. OpenAI uses codex login, Gemini uses Google OAuth, and Qwen supports OAuth or an API key. DeepSeek is supplied through the separately referenced deepseek-mcp-server package and requires its own API key. The Qwen wrapper can use the QWEN_AUTH_TYPE environment variable to select its authentication mode. Docker is also documented, but local provider credentials must be mounted into the container.

Tools and capabilities

The three included servers expose these tools:

• OpenAI: openai_chat and openai_review
• Gemini: gemini_chat and gemini_analyze
• Qwen: qwen_chat

The spyrae/claude-concilium MCP server supports parallel calls, provider-specific error detection, fallback handling, optional iteration on disagreements, and standalone use of each included server. The repository also documents templates and guidance for adding another LLM provider, changing the fallback chain, and customizing prompt strategies.

Limitations and notes

This repository wraps external command-line programs rather than directly implementing the provider APIs. The codex, gemini, and qwen CLIs must be installed and authenticated for their corresponding servers to return provider responses. Primary provider access is described as using OAuth, while Qwen can use an API key and DeepSeek uses a pay-per-use API key.

The included MCP servers are designed around Claude Code usage. The README does not document a hosted endpoint or a single package-manager command that launches the entire project; it describes cloning the repository and installing dependencies inside each server directory. Individual servers can be selected instead of deploying all three.