Cross-examine your AI before you commit

What it does

claude-council is a Claude Code plugin that fires the same prompt to multiple LLM providers—Gemini, OpenAI, Grok, Perplexity, or local CLI equivalents—and renders their answers side-by-side in a tmux pane. It then synthesizes consensus, flags genuine disagreements, and can assign specialized roles like security or devil's advocate to each model. The goal is to make high-stakes decisions—architecture calls, framework picks, security reviews—less dependent on any single vendor’s priors.

The interesting bit

The plugin doesn’t just aggregate raw outputs; its --agents mode spawns parallel Claude subagents to independently evaluate each provider’s response for quality, retrying vague answers and surfacing blind spots before synthesis. There’s also a --debate mode that runs a second round where models critique each other, and an admirably honest --local mode that convenes a council of Claude subagents with different roles while admitting they all share the same training data.

Key highlights

• Queries multiple providers simultaneously with vendor-colored banners in a streaming tmux pane
• Supports CLI-based authentication for Codex, Antigravity, and Grok as an alternative to API keys
• Assigns specialized roles or presets like balanced and architecture, plus a --debate mode for multi-round rebuttals
• --agents mode adds a meta-analysis layer where subagents grade responses and follow up before final synthesis
• Background async jobs for long-running queries, plus an optional stop-gate that reviews uncommitted diffs

Caveats

• Agent-enhanced mode adds roughly 15–25 seconds of latency and ~4× more Claude API usage than standard mode
• The local council mode admits its “members” are all Claude instances, so agreement reflects shared priors rather than independent corroboration
• The README truncates during the auto-context section, leaving the full scope of automatic file detection unclear

Verdict

Worth a look if you already use Claude Code and want to pressure-test high-stakes decisions against model bias. Skip it if you need a single fast answer or if paying for multiple provider APIs feels like overkill.