automagik-genie

[!IMPORTANT] Genie's npm distribution is being soft-deprecated in favor of a cosign + SLSA-verified install through our own CDN.

The canonical install going forward is:

curl -fsSL https://get.automagik.dev/genie | bash

Existing operators on npm install -g @automagik/genie continue to work — the npm package will become a thin postinstall shim that downloads and runs the verified installer. See Security & Trust → Distribution Sovereignty for the threat model and verification flow.

🚀 267 commits this week · 0 releases · +33.1K LoC · 5 contributors

📊 Full velocity dashboard →

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

Manual install (canonical pgserve first)

If you'd rather lay the pieces down yourself, install canonical pgserve before genie. Genie is a consumer of the pm2-supervised pgserve daemon — it never spawns its own. See docs/install.md for the rationale and migration tips.

# 1. Canonical pgserve (registers a pm2-supervised singleton)
bun add -g pgserve@^2
pgserve install

# 2. Genie (will refuse to install if pgserve is unreachable)
bun add -g @automagik/genie
genie install

# 3. Verify
genie doctor

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 genie-spawned agents do — OTel-derived tool, cost, and API-request rows only cover sessions launched via genie spawn / genie team create; user-initiated Claude Code sessions are captured only when they export the OTLP env themselves.

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:

What's new in v4

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

Full v4 release notes →

Observability substrate (v0)

Structured event emission, signed subscription tokens, 4-tier retention, and an external dead-man's switch. Consumer-compat promises are published at docs/observability-contract.md; the 5-phase rollout plan at docs/observability-rollout.md.

Executor read endpoint

External consumers (e.g. the omni scope-enforcer) can query ground-truth turn state directly from genie serve. Two paths are supported:

HTTP — GET http://127.0.0.1:<port>/executors/:id/state returns {state, outcome, closed_at} as JSON. The default port is pgserve_port + 2; override with GENIE_EXECUTOR_READ_PORT. Returns 404 for unknown IDs, 400 for non-UUID IDs, 200 otherwise. No authz — executor IDs are random UUIDs and the view exposes no secrets.

Direct SQL — connect to genie-PG as the read-only executors_reader role and SELECT state, outcome, closed_at FROM executors_public_state WHERE id = $1. Layer login credentials on top:

CREATE ROLE omni_scope_enforcer LOGIN PASSWORD '…' IN ROLE executors_reader;

Response shape (state / outcome / closed_at) is the stable boundary contract; removing or renaming fields is a coordinated breaking change.

Design

Genie ships a single, dark-only color palette inspired by Severance — the Lumon MDR room. One source of truth (packages/genie-tokens/), three consumers (TUI, desktop app, tmux). See docs/design-system.md for tokens, the Severance rationale, how to add a color, how to regenerate the tmux theme, and the visual snapshot workflow.