Claude Concilium is a multi-agent AI consultation framework used to give Claude Code access to the perspectives of other large models through MCP (Model Context Protocol). Its core idea is: although Claude Code is powerful, a single model may miss flaws, overlook edge cases, or get stuck in local optima, and key decisions need diverse perspectives. Concilium uses the standard MCP protocol to let multiple LLMs be consulted in parallel. Each LLM server wraps a CLI tool, and the main providers do not require API keys (they use OAuth authentication).
In terms of main capabilities: it supports parallel consultation with 2 or more AI agents; it provides a production-grade fallback chain with error detection; each MCP server can run independently or be used as part of Concilium; it uses a plug-and-play approach, where you clone the repository, run npm install, and add it to .mcp.json; optionally, after installing the skill, you can use /ai-concilium in Claude Code to trigger multi-agent consultation.
The usage flow includes: formulate the question (a concise description, within 500 characters); send in parallel (OpenAI and Gemini receive the same prompt); handle errors (if a provider fails, the fallback chain starts, for example Qwen → DeepSeek); synthesize (compare responses and look for consensus); optionally iterate (ask follow-up questions about disagreements); make a decision (apply the synthesized solution). All servers detect provider-specific errors and return structured responses.
Installation and configuration steps: 1. Clone and install; 2. Choose at least 2 providers for setup; 3. Add to Claude Code, either by copying config/mcp.json.example and updating the paths, or by adding the servers individually to an existing .mcp.json; 4. Optionally install the skill by copying the Concilium skill to the Claude Code commands directory, after which you can use /ai-concilium in Claude Code to trigger multi-agent consultation. Each server can be used independently, and it is not necessary to install all of them. DeepSeek uses the existing deepseek-mcp-server npm package and does not require a custom server.
Applicable scenarios: when Claude Code alone is not enough, obtain the second and thirdopinions from other LLMs, for parallel code review and problem solving, especially suitable for key decisions that require diverse perspectives.
Limitations and notes: the CLI tools wrapped by the servers (codex, gemini, qwen) require local authentication; when running any server in a container, authentication credentials need to be mounted. For customization, refer to docs/customization.md, which covers adding your own LLM provider, modifying the fallback chain, MCP server templates, custom prompting strategies, and more. Related documentation also includes architecture notes, setup guides for OpenAI/Gemini/Qwen/DeepSeek, and more. Version history: v2.0.0 (2026-03-02) introduced passing prompts via stdin (-p -), OAuth authentication type support via the QWEN_AUTH_TYPE environment variable, added AUTH_NOT_CONFIGURED error detection, graceful shutdown handling (SIGTERM), and increased the default timeout from 90 seconds to 180 seconds; v0.1.0 (2025-12-15) initial version included 3 MCP servers (OpenAI, Gemini, Qwen), the Concilium skill with a fallback chain, and a smoke test suite.
