claude-consensus

Multi-model code review, plan review, and general review for Claude Code and Codex

Multiple AI models independently review code, implementation plans, documents, or technical proposals, then converge through structured synthesis and approval rounds. The lead host participates directly: Claude leads in Claude Code, and Codex leads in Codex.

Quick Start

Claude Code

Tell Claude:

"Set up claude-consensus: https://github.com/AltimateAI/claude-consensus"

Manual install:

/plugin marketplace add AltimateAI/claude-consensus
/plugin install consensus
/consensus-setup

Then use:

/code-review "staged changes"
/plan-review "Add caching to the API layer"
/consensus:review docs/architecture.md

Codex

From a local clone, add this repo marketplace to Codex and install the consensus plugin:

git clone https://github.com/AltimateAI/claude-consensus.git

Use the repo marketplace at:

.agents/plugins/marketplace.json

Then configure and run:

consensus-setup
consensus-code-review staged changes
consensus-plan-review Add caching to the API layer
consensus-review docs/architecture.md

Alias skills are also provided for code-review, plan-review, plan-reviwe, and claude-consensus.

Prerequisites

Minimal setup: the host lead plus 1 external model is enough for consensus reviews.

Supported Panels

Codex intentionally excludes Codex/GPT from the external panel because Codex is already the lead reviewer.

Configuration

Claude setup can smoke-test live model calls. Codex setup does not run live smoke tests unless you ask, because they consume credits.

Example Codex config:

{
  "version": "1.7.0",
  "lead": {
    "id": "codex",
    "name": "Codex",
    "excluded_model_ids": ["gpt", "codex"]
  },
  "models": [
    {
      "id": "gemini",
      "name": "Gemini via Antigravity",
      "command": "agy",
      "resume_flag": "",
      "enabled": true
    },
    {
      "id": "kimi",
      "name": "Kimi K2.6",
      "command": "kilo run -m openrouter/moonshotai/kimi-k2.6 --auto",
      "resume_flag": "-c",
      "enabled": true
    }
  ],
  "min_quorum": 3
}

Commands And Skills

How It Works

Host lead  Model 1  Model 2  Model N
    |         |        |        |
    v         v        v        v
Independent review or planning
    |_________|________|________|
              |
              v
Synthesis: consensus, conflicts, comparison table
              |
              v
Convergence: APPROVE / CHANGES NEEDED, max 2 rounds
              |
              v
Final result with attribution

• Quorum: configurable per host.
• Graceful degradation: unavailable models are skipped if quorum still holds.
• Session artifacts: saved in /tmp/ for debugging and auditability.
• Independent reviews: each model starts from the same prompt with no cross-contamination.
• Codex isolation: Codex uses separate config and never writes Claude user config.

Contributing

Contributions are welcome. Please read the Contributing Guide before submitting a PR.

This project follows the Contributor Covenant Code of Conduct.

License

MIT License. Copyright (c) 2026 Altimate AI.