deliberation: Get Multi-Model Reviews and Consensus

Overview

Deliberation connects coding agents to external models so developers can get independent perspectives on plans, code, security, and debugging questions. Rather than simply returning several raw answers, it synthesizes responses and offers a consensus workflow that records an initial verdict, gathers reviews, and resolves objections over multiple rounds.

It is aimed at developers already working with supported agent hosts who want a structured review step for decisions where a missed issue matters. For straightforward tasks, the extra model calls and setup are unlikely to be worthwhile.

Key Features

• Exposes model and expert-review tools through an MCP server, with a Claude Code plugin and native artifacts for several other coding-agent hosts.
• Supports GPT through the Codex CLI, Gemini through the Antigravity CLI, Grok through an API key, and configured OpenRouter models.
• Provides seven specialist roles, including architecture, code review, security analysis, planning, research, and debugging.
• Offers parallel second opinions and a multi-round consensus workflow with peer review, adjudication, and a configurable round limit.
• Separates advisory analysis from implementation, with the choice guided by the request.
• Includes optional run analytics, session history, and a local dashboard for monitoring activity and provider health.

Use Cases

• A developer can ask for an architecture or migration plan to be reviewed before implementation, especially when tradeoffs or edge cases are difficult to assess alone.
• A team member preparing a security-sensitive change can request a focused security review and use the findings to decide what to fix.
• Someone stuck on a failing test after several attempts can seek another model's debugging perspective before changing more code.
• Developers comparing model opinions on a consequential proposal can use consensus to track objections and see whether the reviewers converge.

Getting Started

For Claude Code, install the marketplace plugin and run setup:

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

For other MCP hosts, the standalone server is available as @antonbabenko/deliberation-mcp. See the README for provider prerequisites, host-specific setup, and configuration details.

Alternatives

• AgentBridge: AgentBridge coordinates Claude Code and Codex for cross-review and task splitting, while Deliberation routes questions to external models for second opinions or consensus.

Considerations

At least one provider must be configured. GPT and Gemini rely on their respective command-line tools and logins, while Grok and OpenRouter require API credentials; OpenRouter models also need configuration. Multi-round consensus can use more time and model calls than a one-shot review, and it is intended for issues where that added scrutiny is useful. The dashboard, debug logging, and session features are optional and have separate configuration.