Getting started

From nothing to a narrated demo

Six steps. The first four are setup you do once; the last two are the part you will repeat every time you ship something worth showing.

Download Corgy for Mac

macOS 15 or later · Apple Silicon

Before you start

✓A Mac on Apple Silicon, running macOS 15 or later — or just Chrome, if the whole demo is a browser tab
✓Claude Code, Codex or ChatGPT
✓Whatever your demo happens in — a browser, a terminal, your own app

Create your account

Sign in at corgy.ai with Google, or with a link emailed to you. A new account starts with 1,000 credits, which is enough to plan, narrate and voice a first demo without deciding anything about money.

Credits are only spent on the parts a model does for you — planning a recording, writing the narration, voicing it. Cutting clips and exporting are free, however many times you redo them. See what things cost.

Install a recorder

The Mac app drives a real cursor across your whole desktop — a browser, a terminal, an editor, a native app — which is why a demo of a terminal and a browser can be one video rather than two. The Chrome extension records a tab, a window or the screen from inside the browser, and needs no Mac at all.

Download the disk image Add the Chrome extension

macOS 15 or later · Apple Silicon · the extension needs Chrome 137 or later

Drag Corgy into Applications, open it, and sign in with the account from step one. Then open Settings → Command line → Install, which puts a corgy command on your PATH. Your coding agent needs that name to drive a recording; the app and the command are the same program.

Recording in the browser? The extension is the whole of setup — you write the prompt in its side panel and press Record there. The macOS permissions in step three are the Mac recorder's, so skip them.

Grant the two macOS permissions

Corgy needs Screen Recording, to capture the run, and Accessibility, to move the cursor and read the labels on screen. macOS will ask the first time.

macOS attaches these permissions to whatever launched Corgy, not to Corgy itself. Started from the app, that is Corgy.app. Driven from an agent session, it is your terminal — so that terminal is what has to be ticked in System Settings, and it has to be restarted afterwards for the grant to take. This is the single most common reason a first recording does nothing.

Add Corgy to your AI assistant

Corgy ships as a plugin for Claude Code, Codex and ChatGPT. Add it once; the first time it is used, it asks you to sign in to Corgy.

In Claude Code, run these two commands once, in any project:

>/plugin marketplace add stormkit-io/agent-skills
>/plugin install corgy@stormkit

In Codex or ChatGPT, add the Corgy plugin from stormkit-io/agent-skills. In ChatGPT you can also connect just the editor: Plugins → + → Add custom MCP server, with https://api.corgy.ai/api/mcp.

That adds the demo skill (/corgy:demo in Claude Code), which reads the repository you are standing in — the router, the real URLs, the labels as they are actually rendered — and plans the demo against screens that exist rather than ones it guessed.

Describe the demo, then read the plan

Change into the repository of the thing you want to show, and say what the demo is for in one sentence.

>/corgy:demo "a 60-second onboarding demo for new users"

Back comes the demo as scenes: what each take proves, the screen it starts from, roughly how long it runs. Nothing records yet. Push back — drop a scene, reorder them, ask for one cut instead of four — and when it looks right, say go. Corgy records the takes one at a time, walking each flow once to prove every target resolves before it rolls.

In Codex or ChatGPT there is no slash command: ask in words, like “plan a 60-second onboarding demo for new users”. ChatGPT cannot see your repository, so it plans from your product's website instead.

No Mac yet? The skill still plans the demo and writes the narration prompt without the recorder — you just cannot record from it.

Edit, narrate, export

The takes land in the editor as a timeline that is already cut into clips.

Trim what dragged, freeze a frame where something needs explaining, and let Corgy write the narration — it is written against what the recorder actually clicked, not guessed from pixels. Pick a voice, render, and you have an mp4 and a link to share.

Open the editor

Let the agent make the whole video

/corgy:make in Claude Code, or “make me a demo video” in Codex, does all six steps for you. It plans the demo from your repository, records every take itself in a clean browser window, checks each take, cuts and assembles the video, writes the narration, and hands back a link to review in the editor. It asks before spending credits on the voiceover.

Add the Corgy plugin to your assistant, as in step four.
Install Corgy for Mac, version 0.1.13 or later, and sign in.
Sign in to Corgy when your assistant first asks. The plugin connects the editor for you; it is how the agent edits the video once the takes are in.
Let agents record. Corgy asks once, the first time an agent starts a recording; you can change it later in Settings → Agents.

>/corgy:make "a 60-second onboarding demo for our dashboard"

Making the whole video needs an assistant that can run commands on your Mac: Claude Code or Codex. ChatGPT cannot record. It plans the demo, you record it in Corgy, and it does the editing, narration and publishing from there.

The takes are recorded on your screen, so leave the mouse and keyboard alone while they run. If a demo needs you signed in, open Settings → Agents → Open the stage browser once and sign in there; it remembers.

When something does not record

The run starts and the screen never changes.: The permissions in step three are attached to the wrong application. Tick the one that launched Corgy — usually your terminal — and restart it.
corgy: command not found: The app is installed but its name is not on your PATH. Open Settings → Command line → Install in the Mac app. If you also build Corgy yourself, the two can fight over the same link — the pane says when it points somewhere else.
A step aims at something that is not there.: The plan named a label that does not match what the page renders. Tell the agent the label as it actually appears, in the words on screen, and ask it to re-plan that scene.
It installs and then will not launch.: The disk image is Apple Silicon only. On an Intel Mac it looks like a broken build, but the machine is the problem.

Still stuck, or want to see how we use it ourselves? Read the blog.