MCP-Based Multi-Agent Code Review Workflow Released

A 4-agent collaborative code review workflow built on the heym platform has been released as open source. This workflow is designed to allow users to self-host it via Docker and maintain security, without sending code to external SaaS (CodeRabbit, Greptile, etc.).

Key Architecture and Features:

• 4-Agent Structure: Consists of 1 Architect, who has no tools and is only responsible for mediation, and 3 Specialists, who use different models (Anthropic, Google, Alibaba, Zhipu) and cognitive frameworks.
• Strict Role Separation: The Architect cannot raise issues on its own and must render the final Verdict based solely on the evidence provided by the Specialists.
• MCP (Model Context Protocol) Support: By wrapping the workflow as an MCP server, Claude Code can automatically invoke it to request a review after completing a task.
• Structured Output: The final output includes a VERDICT, CHANGE_CLASSIFICATION, CONCERNS with supporting evidence, verification tests, and more.

This workflow implements the cognitive scaffolding of the specialist agents using Ejentum's harness API, and is provided under the MIT license.