automagik-dev/genie
Wishes in, PRs out. CLI agent that interviews you, plans the work, dispatches parallel agents in isolated worktrees, and reviews code before you see it.
What it solves
Genie is a planning-and-execution layer for AI coding agents (such as Claude Code and Codex), designed to move from a vague idea to a merge-ready pull request through a structured, multi-step workflow. It prevents the "hallucination loop" by enforcing a strict sequence of brainstorming, design review, planning, and independent implementation review.
How it works
Genie uses a combination of git-based documentation and a local SQLite database to track state. The process follows a specific lifecycle:
- Brainstorming: An idea is refined into a
DESIGN.mdfile. - Wishing: The design is converted into a scoped
WISH.mdwith execution groups. - Planning: The plan is reviewed and approved.
- Work: Native role agents are dispatched to build the approved groups in parallel.
- Review: An independent reviewer agent (separate from the engineer agent) verifies the implementation against acceptance criteria.
Operational state is stored in a single .genie/genie.db file, allowing multiple git worktrees to share the same task state without a background daemon.
Who it’s for
Developers using AI coding agents who want a more rigorous, structured approach to feature development and a way to manage complex tasks that require planning and parallel execution across multiple agents.
Highlights
- Git-native state: Wishes and designs are plain markdown files versioned in git.
- Zero-daemon architecture: No background processes or heavy databases (like Postgres) are required; it uses a SQLite file.
- Warp Cockpit: Can launch a wish's ready groups into a Warp terminal window with one pane per group in its own worktree.
- Omni Bridge: Integrates with WhatsApp via an Omni hub for remote approvals and short tasks.
- MCP Server: Provides a read-only Model Context Protocol (MCP) server to expose task and board state to other AI tools.