Wishes in, PRs out.

Genie is a CLI that turns one sentence into a finished pull request. You describe what you want — Genie interviews you, writes a plan, spawns parallel agents in isolated worktrees, reviews the code, and opens a PR. You approve. You merge. That's it.

Get started

Paste this into Claude Code, Codex, or any AI coding agent:

Install Genie and set up this project:

curl -fsSL https://raw.githubusercontent.com/automagik-dev/genie/main/install.sh | bash
genie
/wizard

That's it. The wizard interviews you, scaffolds your project, and walks you through your first wish. Relax.

Or install manually:

curl -fsSL https://raw.githubusercontent.com/automagik-dev/genie/main/install.sh | bash

What you get

  "Add dark mode"
       |
   /brainstorm ──── Genie asks questions until the idea is concrete
       |
   /wish ────────── Turns it into a plan: scope, criteria, task groups
       |
   /work ────────── Agents spawn in parallel worktrees, each on its own branch
       |
   /review ──────── 10 critics review. Severity-tagged. Nothing ships dirty.
       |
   Pull Request ─── You approve. You merge. Ship it.

Parallel agents. Not one agent doing everything sequentially — multiple agents working at the same time in isolated worktrees. No conflicts. No re-explaining context.

Automated review. A 10-critic council (architecture, security, DX, performance, ops, and more) reviews every design before you see it. Severity-tagged findings. CRITICAL blocks the PR.

Overnight mode. /dream — queue wishes before bed. Wake up to reviewed PRs.

Kanban boards. Built-in task boards with custom pipelines and stage gates.

Postgres-backed. All state lives in PostgreSQL — agents, tasks, events, messages. Queryable. Durable. Real-time via LISTEN/NOTIFY.

Full observability. Events, metrics, session replay, cost tracking. See everything your agents do.

Portable context. Identity, skills, memory — all markdown files in your repo, git-versioned. You own everything.

Any AI provider. Claude, Codex, or any OpenAI-compatible model. Bring your own agent.

Why Genie?

Without Genie Re-explain context every new chat One agent, one file at a time Copy-paste PR descriptions manually Review AI code yourself, line by line Context lost mid-conversation No memory between sessions

With Genie Context captured once, inherited by every agent Parallel agents in isolated worktrees Automated severity-gated code review Queue wishes overnight, wake up to PRs 10-critic council catches what you'd miss Persistent memory across sessions

Skills

17 built-in skills that compose into workflows:

Skill	What it does
/brainstorm	Explore vague ideas with guided questions
/wish	Turn an idea into a scoped plan with acceptance criteria
/work	Execute a wish with parallel agents
/review	Severity-gated code review (SHIP or FIX-FIRST)
/council	10-perspective architectural deliberation
/dream	Batch-execute wishes overnight
/trace	Investigate bugs — reproduce, isolate, root cause
/fix	Minimal targeted bug fixes
/report	Deep investigation with browser + trace
/refine	Transform rough prompts into structured specs
/learn	Correct agent behavior from mistakes
/docs	Audit and generate documentation
/pm	Full project management playbook
/brain	Knowledge graph with Obsidian-compatible vaults
/genie	Auto-router — natural language to the right skill
/genie-hacks	Community patterns and real-world workflows
/wizard	Guided first-run onboarding

What's new in v4

v4 is a ground-up rewrite. 700 commits. 300 files. ~19K lines.

	v3	v4
State	JSON files + NATS	PostgreSQL + LISTEN/NOTIFY
UI	CLI only	Full terminal UI
Memory	None	Optional knowledge brain
Tasks	Basic	Kanban boards, templates, projects
Observability	Minimal	OTLP, session capture, audit trail
Review	Single pass	10-critic council deliberation
Stability	Best effort	Advisory locks, spawn watchdog, 200+ bug fixes

Full v4 release notes →

---

Docs · v4 Release · Discord · MIT License

_{You describe the problem. Genie does everything else.}