Claude Council: Multi-Model Review and Debate Plugins Compared

What a "Claude Council" Search Actually Finds

Search claude council and what you find is a pattern, not a single product: Claude Code plugins that run several models — or several differently prompted agents — in parallel on the same question, then synthesize the answers into one structured response. The most installed plugin in the cluster, llm-council-plugin, pulled 80 installs in the past 7 days by orchestrating Claude, OpenAI Codex, and Gemini into parallel code reviews that end in a consensus verdict. This post compares eight council-style plugins on what they actually ship — commands, agents, skills, hooks — and where each fits, so you can decide whether you want a multi-model review council, an adversarial debate setup, or a perspective council built from role prompts.

There is a plugin literally named claude-council (hex-claude-council, 269 stars), but it sits in a cluster of plugins implementing the same idea under names like llm-council, agent-council, polyclaude, and design-council. In a directory of 37,991 plugins this is a small category, and weekly install numbers are modest across the board — but the differences between the plugins are concrete: some call other providers' models by name, others build the council from role-specialized perspectives, and they ship very different component mixes.

Claude Council Plugins to Consider

llm-council-plugin (xrf9268-hue-llm-council-plugin) orchestrates multi-model code reviews: it runs Claude, OpenAI Codex, and Gemini in parallel on the same code, then synthesizes their opinions into a consensus verdict, with automated setup and cleanup. It ships commands, agents, skills, and hooks — the hooks carry the setup and cleanup automation. At 80 installs over the past 7 days (17 stars), it is the most installed plugin in this set. Use it when you want disagreement between models surfaced on a diff before you merge it.

llm-council (sherifkozman-llm-council) runs adversarial debates between multiple LLMs to cross-validate code, architecture, reviews, security, research, and planning tasks, producing specialized outputs like code, reviews, or plans. It ships commands and skills (9 installs/7d, 70 stars). Where llm-council-plugin is scoped to code review with a consensus goal, this one is built around structured disagreement across a wider range of task types — reach for it when the thing you are validating is a plan or an architecture, not a diff.

claude-council (hex-claude-council) is the plugin the query names. It queries Gemini, OpenAI, Grok, Perplexity, and Claude in parallel for diverse perspectives on architecture decisions, technology choices, and debugging dead-ends, then returns a structured synthesis of consensus and divergence. It ships commands, agents, and skills, and at 269 stars (18 installs/7d) it has the highest star count among the dedicated council plugins here. Its provider list is the widest in this set — including Grok and Perplexity — so use it when you are stuck on a decision and want the broadest spread of independent opinions.

pro-workflow (rohitg00-pro-workflow) is not a standalone council: multi-LLM council deliberation is one feature inside a complete AI-assisted coding workflow with self-correcting memory, persistent FTS5-indexed research wikis, auto-research loops, and 8 specialized agents that coordinate parallel sessions, enforce quality gates, audit context costs, and capture learnings across sessions. It ships commands, agents, skills, and hooks. At 2,269 stars it is the most starred plugin in this set, though installs are low (2/7d). Pick it if you want council deliberation embedded in a full workflow system rather than added to your existing one.

llm-council (dair-ai-llm-council-plugins-llm-council) implements Karpathy's LLM Council pattern with open-weight models on Fireworks AI: models generate individual responses, rank their peers' outputs, and a Chairman synthesizes the consensus answer. It is skills-only, at 1 install over the past 7 days (its source repo hosts 2 plugins, so repo stars are not attributable to it). It is the only plugin in this set targeting open-weight models — relevant if your council should not depend on closed-API providers.

agent-council (team-attention-agent-council-plugins-agent-council) queries a council of multiple AI agents for diverse perspectives on code tasks or decisions, then returns a synthesized response that reconciles their opinions. It is also skills-only, with one distinctive behavior in this set: it can trigger automatically when it detects a multi-perspective request, in addition to manual invocation via the /agent-council command, using bash-orchestrated agent interactions. (Its source repo hosts 13 plugins, so star counts are not attributable.) Use it if you would rather the council convene itself than remember to call it.

polyclaude (riley-coyote-polyclaude) takes the perspective approach: it spawns a configurable council of viewpoints — User Advocate, Architect, Skeptic — to analyze questions, plans, or ideas from multiple angles in parallel, generating structured reports with synthesized verdicts and key tensions. It ships commands and skills (175 stars). Unlike hex-claude-council or llm-council-plugin, the diversity comes from configured perspectives rather than separate model providers — the lens varies instead of the model.

design-council (sjsyrek-design-council) convenes parallel role-specialized peer agents — principal engineer, platform, integration, test, QA, security — to debate cross-domain technical decisions or audit codebases in real time, with Claude acting as CEO: routing peer DMs, arbitrating deadlocks, and producing a one-page decision log. It ships commands and skills (151 stars). Compared to polyclaude's general-purpose perspectives, its roles are engineering-specific and the session ends with a written decision artifact — suited to decisions you need to document, not just make.

How They Differ: Component Mix and Use Cases

The clearest split is where the council's diversity comes from. One group calls external model providers by name: llm-council-plugin (Claude, OpenAI Codex, Gemini), hex-claude-council (Gemini, OpenAI, Grok, Perplexity, Claude), sherifkozman's llm-council (multiple LLMs in adversarial debate), and dair-ai's llm-council (open-weight models on Fireworks AI). The other group builds the council from role prompts and peer agents: polyclaude's perspectives, design-council's engineering roles with Claude as CEO, and agent-council's bash-orchestrated agents. External-model councils give you opinions from genuinely independent models; perspective councils vary the lens instead of the model.

Component mix is the second axis — and a reasonable proxy for how much lifecycle automation a plugin carries (see Claude Code plugin examples by component type for what each component type does in practice). llm-council-plugin and pro-workflow ship commands, agents, skills, and hooks; the hooks are what let llm-council-plugin automate review setup and cleanup, and pro-workflow enforce quality gates across sessions — the same hook mechanism covered in automating Claude Code workflows. At the other end, dair-ai's llm-council and agent-council are skills-only: there is no hook-driven automation, and invocation happens through skill activation (automatic, in agent-council's case). hex-claude-council, sherifkozman's llm-council, polyclaude, and design-council sit between, pairing commands with skills (plus agents, for hex-claude-council).

The third axis is scope. llm-council-plugin is purpose-built for code review; sherifkozman's llm-council and hex-claude-council target a broader range of decisions — architecture, security, research, technology choices, debugging dead-ends; design-council narrows to cross-domain engineering decisions and codebase audits; and pro-workflow treats the council as one stage in an end-to-end workflow. Note the metrics tension in this set: pro-workflow has by far the most stars (2,269) but only 2 installs in the past 7 days, while llm-council-plugin has just 17 stars but 80 weekly installs — star count and current install activity point at different plugins here.

Choosing One for Your Workflow

• Pre-merge review with multiple models: llm-council-plugin — parallel Claude/Codex/Gemini review with a consensus verdict and hook-automated setup.
• Cross-validating a plan or architecture: sherifkozman's llm-council — adversarial debate across code, architecture, security, research, and planning tasks.
• Stuck on a decision or debugging dead-end: hex-claude-council — the widest provider spread (Gemini, OpenAI, Grok, Perplexity, Claude) with consensus and divergence called out separately.
• A full workflow system that includes a council: pro-workflow — council deliberation plus memory, research wikis, and 8 specialized agents.
• Open-weight models only: dair-ai's llm-council — the Karpathy LLM Council pattern on Fireworks AI.
• Hands-off invocation: agent-council — auto-triggers on multi-perspective requests.
• Multiple viewpoints without external providers: polyclaude — configurable perspective councils with structured reports.
• Decisions that need a paper trail: design-council — engineering-role debate that ends in a one-page decision log.

Wrap-up

The council category is small and its install numbers are honest about that — the largest figure in this set is 80 installs over 7 days. But the pattern is well-defined, and the eight plugins here differ on axes you can actually evaluate: external models versus configured perspectives, hooks-driven automation versus skills-only invocation, and code-review focus versus general decision support. Start from the axis that matters to you, check the plugin's component mix against it, and browse the wider plugin directory if your workflow needs something the council pattern does not cover.