Skip to content

CADENCE User Guide

CADENCE is a draft/build/settle framework for AI-assisted development that brings GSD-grade discipline to iterative work with configurable quality gates at each phase.

Cadence has one core engine and three surface categories — CLI, host adapters, and an MCP server — for four current entry points, not separate tools:

  • The cadence CLI is the engine — it implements the DRAFT→BUILD→SETTLE loop and all quality gates. You run it in a terminal; a human operator or an AI agent can drive it. Completely host-agnostic and model-agnostic.
  • Host adapters wire the same engine into a specific coding agent — currently two:
    • The cadence-host-claude-code install adapter is a thin wiring layer that plugs the same engine into Claude Code via lifecycle hooks and 15 slash commands. Claude Code is the reference adapter for ambient edit-time gates (boundary checks and anomaly detection as you edit).
    • The cadence-host-codex install adapter plugs the same engine into the OpenAI Codex CLI via lifecycle hooks and global prompt commands. It is the second shipped conformance consumer of the host-adapter contract, and also supports ambient edit-time boundary checks via its own hooks.
  • cadence mcp serve runs the engine as a local MCP server over stdio, so any MCP-capable host (Claude Desktop, Cursor, other agents) can drive the loop with no bespoke adapter. It exposes the imperative loop only — command-boundary gates run; ambient edit-time gates require host hooks.

Read in this order:

  • Quickstart — Try the no-install demo, start a real phase from a template, or walk the full loop end-to-end.
  • Concepts — Understand the loop, gates, profiles, and single-commit convention that everything else builds on.
  • CLI guide — Master the cadence command-line tool: all subcommands, flags, and workflows.
  • Claude Code integration — Set up and use the Claude Code slash commands and understand how the adapter bridges the CLI engine.
  • Host adapters - Understand the shared adapter contract used by Claude Code and Codex.
  • MCP server — Drive the loop from any MCP host (Claude Desktop, Cursor, agents) via cadence mcp serve.
  • Providers — Integrate with your model provider: mock (offline default), anthropic, local (OpenAI-compatible endpoints), or host-cli.
  • Command reference — Exhaustive reference for all CLI subcommands and their options.
  • Config reference — Full .cadence/ config schema and all settable fields.
  • Release process — Maintainer checklist for publishing npm, tag, and GitHub Release records in sync.

Start with Quickstart to see the loop in action, then move to Concepts to understand why each phase exists.