claude-council

A Claude Code plugin that consults multiple AI coding agents in parallel and shows you their answers side-by-side. Useful when one model's bias could mislead you and the right call depends on cross-checking — architecture decisions, debugging dead ends, security reviews, framework picks.

Five providers answering the same question. Each banner names the provider, the model that answered and how long it took. The synthesis separates what they all agreed on from where they diverged — here, whether trapping EXIT INT TERM on one handler is correct, or whether signals should be converted into exits first — and when they agree instead, it names the assumption the answer rests on.

Quick start · Usage · Configuration · Reference · Development

Quick start

# 1. Install via Claude Code plugin marketplace
/plugin marketplace add hex/claude-marketplace
/plugin install claude-council

# 2. Configure at least one provider — any of these works:
export OPENAI_API_KEY="..."         # or GEMINI_API_KEY, XAI_API_KEY, PERPLEXITY_API_KEY, KIMI_API_KEY,
                                    # OPENROUTER_API_KEY
                                    # OR install the codex / antigravity (agy) / grok / kimi CLIs (uses your
                                    # existing subscription — no API key needed)

# 3. Ask anything
/claude-council:ask "Should I use UUID or BIGINT primary keys for a SaaS users table?"

You get side-by-side responses from each configured provider:

🔳 Codex - default
   Use UUID primary keys — they avoid enumeration, work across distributed
   services, and survive imports/exports cleanly.

🟦 Antigravity - default
   UUIDv7 specifically: security of non-guessable IDs plus the index
   locality of time-ordered sequences.

🟥 Grok - grok-latest
   BIGINT autoincrement — smaller index, faster joins. Handle public-
   exposure concerns with a separate UUID slug column.

🟩 Perplexity - sonar-reasoning-pro
   BIGINT: 25% smaller than UUID, better cache locality, with citations
   to Postgres benchmarks.

## Synthesis
Two providers prefer UUID(v7), two prefer BIGINT. Choice depends on
whether you need distributed ID generation.

When they all agree instead, the synthesis says what that agreement rests on, because agreement is where it is easiest to stop asking:

## Synthesis
All five recommend SQLite. Read that as agreement about the reasoning, not
as verification: every provider was given the same description of a system
none of them can inspect. The answer assumes this stays single-node — the
one premise that would flip it, and the one nobody here could check.

Inside tmux, results stream into a side pane in real time with vendor-colored banners. Run /claude-council:status to confirm what's configured and connected.

Features

• Query Gemini, OpenAI (GPT/Codex), Grok, Perplexity, and Kimi (Moonshot AI) simultaneously
• Seat any model OpenRouter routes to — Anthropic's Claude by default, so the council hears the one vendor it otherwise has no voice for
• Use the codex, agy (Antigravity), grok, and kimi (Kimi Code) CLIs (subscription auth) when installed — preferred over their API siblings
• Run a local ollama model as a council member — no key, no subscription, no network
• Side-by-side comparison of responses with vendor-colored headers
• Streaming tmux pane that renders responses as they land
• Specialized roles, debate mode, and agent-enhanced deep analysis for high-stakes decisions
• Background jobs (--async) for long-running queries, with /claude-council:result to fetch, list, and cancel
• Opt-in stop-gate: a second model reviews your uncommitted diff before Claude ends its turn
• Extensible provider system — add new AI agents easily
• Put the conversation itself to the council with /claude-council:advise, which shows you what would leave the machine before it goes
• Proactive agent that suggests consulting the council on architecture / debugging dead ends

Installation

From Marketplace (Recommended)

# Add the hex-plugins marketplace
/plugin marketplace add hex/claude-marketplace

# Install claude-council
/plugin install claude-council

Manual (run from a local clone)

For normal use, prefer the marketplace or GitHub install above — both persist across sessions. A manual clone is for running from a local working copy (development, or offline). Clone the repo anywhere, then point Claude Code at the repo root for the current session:

git clone https://github.com/hex/claude-council.git
claude --plugin-dir /path/to/claude-council    # repo root; loaded for this session only

Cloned it and nothing loads? Two traps to avoid:

1. Don't clone into ~/.claude/plugins/ (Windows: %USERPROFILE%\.claude\plugins\). That's Claude Code's managed install cache — it is never scanned for manually-added plugins, so the plugin won't appear in the Installed tab or respond to its slash commands.
2. pluginDirectories in settings.json does nothing — it isn't a real setting, so it's silently ignored (no error shown). Use --plugin-dir above for a local clone, or install via the marketplace / GitHub for a persistent setup.

Usage

Slash Commands

# Query all configured providers
/claude-council:ask "How should I structure authentication in this Express app?"

# Query specific providers
/claude-council:ask --providers=gemini,openai "What's the best approach for caching here?"

# Include a specific file for review
/claude-council:ask --file=src/auth.ts "What's wrong with this implementation?"

# Attach a screenshot for visual critique
/claude-council:ask --image=shot.png "Why does this dialog render off-center?"

# Export response to markdown file
/claude-council:ask --output=docs/auth-decision.md "How should we implement authentication?"

# Quiet mode - show only synthesis
/claude-council:ask --quiet "What's the best caching strategy?"

# Check connectivity and configured models for each provider
/claude-council:status

# Run a long query in the background, fetch it later
/claude-council:ask --async "Deep-dive the tradeoffs of event sourcing here"
/claude-council:result <job-id>

Quick Reference

Specialized Roles

Assign different perspectives to each provider for more comprehensive reviews:

# Use specific roles
/claude-council:ask --roles=security,performance,maintainability "Review this auth code"

# Use a preset
/claude-council:ask --roles=balanced "Review this implementation"

# Bind a role to a named provider instead of a position
/claude-council:ask --roles=openrouter-2=security,perplexity=devil "Review this design"

Available roles:

• security - Security Auditor (vulnerabilities, OWASP Top 10)
• performance - Performance Optimizer (efficiency, bottlenecks)
• maintainability - Maintainability Advocate (clarity, future changes)
• devil - Devil's Advocate (challenges assumptions)
• simplicity - Simplicity Champion (identifies over-engineering)
• scalability - Scalability Architect (growth, scaling)
• dx - Developer Experience (API ergonomics)
• compliance - Compliance Officer (GDPR, regulations)

Presets:

• balanced - security, performance, maintainability
• security-focused - security, devil, compliance
• architecture - scalability, maintainability, simplicity
• review - security, maintainability, dx

A bare list is positional: the first role goes to the first provider discovery returns, the second to the second, and so on. That is fine for a fixed roster and fragile for a growing one — adding a provider script shifts every later provider's role by one, and reordering OPENROUTER_MODELS reassigns which router seat plays which part. Both happen silently, because only non-empty roles are printed.

provider=role pairs bind the two explicitly and survive both. The two forms cannot be mixed in one --roles (a bare entry alongside a keyed one is ambiguous); a pair naming a provider that is not being queried, or naming one provider twice, is refused rather than resolved; and any provider left without a role is named on stderr:

Note: no role for openai grok

Debate Mode

Enable multi-round discussions where providers critique each other:

/claude-council:ask --debate "How should I structure the database schema?"

How it works:

1. Round 1: All providers answer the question normally
2. Round 2: Each provider sees the others' responses and provides rebuttals
3. Synthesis: Incorporates debate insights, consensus shifts, and unresolved tensions

Debate mode surfaces blind spots and stress-tests recommendations. The synthes