Deliberation: Multi-Agent LLM Consensus for Code and Plan Review

Summary

Deliberation is an innovative GitHub repository that enables Claude Code to leverage multiple LLMs like GPT, Gemini, Grok, and 400+ OpenRouter models for expert second opinions and arbiter-mediated consensus. It provides specialized AI agents for tasks such as code review, security analysis, and architectural design, ensuring comprehensive and reliable feedback. This project helps developers get diverse perspectives and achieve higher quality in their work.

Use at your own risk

OSRepos shares public repositories for knowledge and discovery only. Any installation, execution, configuration, or use of code from these repositories is the user's own responsibility. Always review the repository, source code, dependencies, licenses, and security implications before running or installing anything. OSRepos is not responsible for issues, damages, or losses resulting from third-party repositories.

Introduction

Deliberation is an innovative GitHub project by Anton Babenko designed to enhance code and plan reviews by leveraging the power of multiple Large Language Models (LLMs). It integrates with environments like Claude Code, Cursor, VS Code, and Kiro, allowing developers to solicit second opinions and achieve consensus from a diverse panel of AI experts. This system supports models such as GPT, Gemini, Grok, and over 400 models available through OpenRouter, providing specialized agents for various development tasks.

Installation

Deliberation offers flexible installation methods, primarily as a plugin for Claude Code or as a standalone Multi-Agent Communication Protocol (MCP) server.

For Claude Code plugin (recommended):

/plugin marketplace add antonbabenko/agent-plugins
/plugin install deliberation@antonbabenko
/deliberation:setup

This setup registers the MCP servers, allowing Claude to route complex tasks to your configured LLM experts.

For standalone deliberation MCP server:

The orchestration server is also available as an npm package. You can install it via npx and configure it for various MCP clients like Cursor, VS Code, or Kiro.

npx -y @antonbabenko/deliberation-mcp

Provider prerequisites include installing the Codex CLI for GPT, Antigravity CLI for Gemini, setting XAI_API_KEY for Grok, and OPENROUTER_API_KEY along with a configuration file for OpenRouter models.

Examples

Deliberation provides powerful commands to orchestrate multi-model interactions:

/consensus: This command initiates an arbiter-mediated convergence loop involving GPT, Gemini, Grok, and Claude. It's designed for deep analysis, where models review a plan, debate edge cases, and reach a consensus. The process involves multiple rounds, starting from initial disagreements and progressing towards a unified verdict, ensuring thorough vetting of proposals.
/ask-all: For quicker, parallel insights, /ask-all dispatches a task to GPT, Gemini, Grok, and configured OpenRouter models simultaneously. Claude then synthesizes their independent findings. This is particularly useful for getting diverse perspectives on architectural designs, where models can critique each other's suggestions to reveal hidden issues.

The project also features seven domain-specific experts, including Architect, Code Reviewer, Security Analyst, and Debugger, each tailored to handle specific problem types.

Why Use Deliberation

Deliberation is invaluable for enhancing development workflows by providing intelligent, multi-faceted AI assistance.

When experts help most:

Architecture decisions: Get diverse opinions on system design and tradeoffs.
Stuck debugging: Obtain fresh perspectives after multiple failed attempts.
Pre-implementation validation: Ensure your plans are sound before writing code.
Security concerns: Identify vulnerabilities and harden endpoints.
Code quality: Receive a second opinion on your implementation.

When not to use experts:

Simple file operations (handled directly by Claude).
First attempts at fixes (try yourself first).
Trivial questions (no need for delegation).

The system employs several "guards" to ensure the integrity of the consensus process, such as blind verdicts from Claude, peer review among external models, and a strict "no self-approval" rule, preventing biased outcomes.