plan-cascade

Plan Cascade is an MCP server that orchestrates AI-powered, multi-agent cascading development across three levels of granularity. It decomposes complex projects into feature stories, auto-generates PRDs and design documents, and executes tasks in parallel using specialized agents like claude-code, codex, aider, and amp-code. You interact with it primarily through the standalone CLI (plan-cascade) or via the Claude Code Plugin integration, letting you run quick tasks or full mega-plans with customizable flow and testing gates. Typical usage involves configuring a plan, choosing an agent strategy, and then triggering execution where the system handles decomposition, design generation, and parallel story execution with built-in quality gates.

Prerequisites:

• Python 3.8+ and pip
• Optional: Git if you plan to use external submodules or skills

Install Plan Cascade from PyPI:

pip install plan-cascade

Configure the tool (example):

plan-cascade config --setup

Run a plan with auto strategy:

plan-cascade run "Implement user authentication with OAuth2" --flow quick

Alternatively, run a mega-plan:

plan-cascade run "Build an e-commerce platform" --flow standard

If you prefer the Claude Code Plugin workflow, install the plugin and use the provided commands described in the Quick Start section of the README.

Tips and common issues:

• Ensure Python 3.8+ is installed and accessible in your PATH.
• If you encounter environment or dependency issues, consider creating a virtual environment: python -m venv venv && source venv/bin/activate (on Unix) or venv\Scripts\activate (on Windows) Then reinstall: pip install plan-cascade
• The server auto-detects framework skills based on your project contents (e.g., React, Vue, Rust) via project files like package.json or Cargo.toml.
• For best results, provide a design document (PRD) or architecture.md to guide mega-plan or specific feature planning.
• When using external agents, you can specify preferred agents for execution or retries (e.g., --impl-agent=aider --retry-agent=codex).
• If you see skeleton code, use the AI Verification Gate to ensure implementations meet acceptance criteria before merging.