AGENTS.md, routes to the right playbook, and calls the ralphy CLI for every model call and render.
This quickstart gets you from a fresh machine to final.mp4 in roughly ten minutes.
The 30-second version
The four steps in order
Install
Clone the repo, open Claude Code, ask Claude to set you up. About 3-5 minutes including grabbing two API keys.
Your first video
Type a brief in chat. Ralphy runs intake, generates scenes, renders
final.mp4. About 8 minutes cold-start.Connect your editor
Using Cursor, Copilot, or Codex instead of Claude Code? One command wires the routing in.
What just happened
Walk through every file Ralphy wrote —
BRIEF.md, prompts.json, assets/, generations.jsonl, final.mp4.What you’ll need
- A terminal on macOS or Linux. Windows works via WSL.
- Claude Code on your machine (Cursor / Copilot / Codex also work — see step 3).
- An OpenRouter key — covers image, video, LLM, vision, transcription.
- An ElevenLabs key — covers voiceover and music.
How Ralphy works in one paragraph
You cloneugc-cli and open it in Claude Code. The repo ships AGENTS.md, a routing table the agent reads on every turn. You write a brief like "15-second TikTok about my espresso bar, morning vibe, selfie POV" in chat. Claude matches the brief to a playbook, asks 3-5 clarifying questions, then calls ralphy for every model invocation and render step. Each project lives under workspace/projects/<id>/ with an append-only generation log, so you can replay, branch, and audit costs after the fact. The agent reviews scenes with you before paid generation; nothing burns budget without consent.
Pick your path
Ifwhich ralphy returns nothing and you haven’t cloned the repo yet, start at Install. If ralphy doctor is already green, skip to “Your first video”.
Related
- What is Ralphy? — the vision in one page
- Architecture — the deeper read after your first render
- Talking to Ralphy — how to phrase briefs and follow-ups
- Install playbook — canonical source
- Prompt library and Skills gallery on the landing