hex/claude-council

Claude Council is a plugin for Anthropic's Claude Code that lets you query several different AI models at once and see their answers side-by-side glong with a synthesis that shows where they agree and where they differ. It's for decisions where a single model's bias could mislead you — architecture choices, debugging dead ends, security reviews, or framework selections.

How It Works

• Parallel queries: You ask a question (e.g. `/claude-council:ask "Should I use UUID or BIGINT primary keys?") and the plugin sends it to all configured providers simultaneously.
• Providers: It can talk to Gemini, OpenAI (GPT/Codex), Grok, Perplexity, Kimi, and any model on OpenRouter. It can also use local CLIs (like Codex, Antigravity, Grok) or a local Ollama model — no API key needed.
• Side-by-side view: In a tmux pane (or in Claude's interface) you see each provider's answer with a colored header naming the provider and model, and finally a synthesis that separates consensus from divergence.
• Synthesis nuance: When all providers agree, the synthesis states the underlying assumption that agreement rests on (e.g. "all were told the same description of a system they can't inspect"). This prevents false confidence from unanimous answers.

Key Features

• Roles – Assign different perspectives (security, performance, devil's advocate, etc.) to each provider via --roles.
• Debate mode – --debate runs two rounds: providers answer, then they see each other's answers and rebut, leading to a deeper synthesis.
• Agent-enhanced analysis – --agents uses Claude analyst agents to critique each provider's response, producing a confidence-weighted consensus and blind-spot analysis (slower and more expensive).
• Local council – --local uses several blind Claude subagents with different roles when you have no external API keys (with an honest caveat that they share training data).
• Auto-context injection – Detects relevant files (e.g. auth → src/auth/*.ts) and includes them in the question.
• Image support – --image attaches a screenshot for vision-capable providers.
• Caching – Responses are cached (keyed by prompt+provider+model+role) to save time and cost.
• Background jobs – --async runs long queries in the background, fetchable later.
• Export – --output saves a markdown decision log.

Installation & Usage

Install via Claude Code's plugin marketplace:

/plugin marketplace add hex/claude-marketplace
/plugin install claude-council

Then set at least one API key (e.g. OPENAI_API_KEY, GEMINI_API_KEY) or install a CLI like Codex. Ask anything with:

/claude-council:ask "Your question"

Or check your setup with /claude-council:status.

Privacy & Caveats

• The full prompt (and any included files) is stored in plaintext in the cache and exported transcripts, though a .gitignore is dropped to prevent committing them.
• Images are sent to vision-capable providers but are not stored in caches or transcripts.
• The local council is honest that multiple Claude subagents share the same training and priors — agreement there isn't cross-vendor verification.

Claude Council is a practical tool for cross-checking AI advice, especially when one model's confident answer could be wrong. It's a real, functional project rather than a wrapper or demo.