LLM-WarRoom: Multi-Model AI Reasoning Workbench and Advisor-Perspective Decision Framework

LLM-WarRoom is a local-first multi-model AI reasoning workbench inspired by Karpathy's LLM Council, positioned as a pragmatic reasoning assistant tool. It supports independent responses, anonymous peer reviews, and provides a pressure-tested decision-making process with five advisor perspectives, helping users obtain multi-model viewpoints, identify points of divergence, and improve the reasoning quality of complex decisions.

As the capabilities of large language models improve, single-model responses can hardly meet the needs of complex decision-making. The LLM Council concept proposed by Andrej Karpathy demonstrates the reliability of multi-model collaboration, and LLM-WarRoom translates this idea into an actionable local tool, with additional inspiration from Ole Lehmann's Claude Council skill.

A local web application using FastAPI for the backend and React for the frontend, following the local-first principle: all conversations and runtime outputs are written to the local data/ folder. The backend includes the FastAPI framework, a model alias system, and multi-provider support; the frontend uses the React tech stack, and the local server is started via npm run dev.

Two Working Modes:

1. Ask Mode: Submit Question → Independent Response → Anonymous Peer Review → Final Synthesis
2. War Room Mode: Frame the Question → Advisor Perspective Analysis → Anonymous Review → Final Decision Five Advisor Perspectives: Contrarian, First Principles Thinker, Expansionist, Outsider, Executor—all are prompt-based roles.

Models are managed via environment variables and configuration files:

• Environment variables need to set API keys (e.g., OPENAI_API_KEY) and a list of model aliases
• Model aliases include openai_primary (gpt-5.1), claude_primary (claude-sonnet-4-20250514), etc.
• Supports OpenAI (direct integration), Anthropic (requires package installation), and OpenRouter (free tier suitable for experiments).

Local-first storage strategy:

• Each runtime output is stored in data/runs//, including run.json (machine-readable), summary.md (human-readable), and decision.json (when applicable)
• Conversation history is stored in data/conversations/, and outputs are ignored by git by default to ensure user data control and audit trails.

Suitable Scenarios: Complex decision analysis, creative divergence and convergence, risk assessment, solution validation Notes: Outputs from this experimental software may be incorrect; prompts are sent to model providers; do not paste sensitive information; keep the .env file private.

LLM-WarRoom represents a new AI tool paradigm, improving reasoning quality through multi-model collaboration. Its local-first design, advisor perspective framework, and flexible configuration make it a powerful tool for complex decisions. Future enhancements can include expanding endpoints (e.g., /api/cases/evaluate) to further enrich functionality.