The Codex workbench
The workbench is a folder you download once and keep. Opened with Codex — inside the ChatGPT desktop app — it turns the agent into your Looky guide: it installs the tools, connects the folder to your instance, and then works with you day to day — pulling workspaces, authoring content, validating and publishing — explaining each step in plain language. No terminal, Python or git knowledge required.
Works on any ChatGPT plan. Codex is included even on the free plan, though its usage allowance is the tightest of the agents Looky supports — a Plus plan gives more room for real sessions. Either way, this workbench is built to live with usage limits: it saves your progress as it goes, so a session that gets interrupted picks up exactly where it left off.
Prefer git? git clone https://github.com/paranoid-software/looky-codex works too, and updates arrive via git pull.
Before you start
- The ChatGPT desktop app, signed in with your ChatGPT account — any plan works.
- Access to your Looky instance: an invitation to an account, as a developer if you'll publish content.
- Windows or macOS — the workbench handles each one's quirks.
Get going in three steps
- Download and extract the ZIP wherever you keep your documents, then rename the folder after your instance — for example
my-lookyformy.looky.studio. - Open the folder in the ChatGPT desktop app. Choose to work in a folder so it can read and write files there, then choose Codex.
- Type
@and pick "Set up the workbench."
The agent asks how you prefer to work and takes it from there. The verification code for signing in arrives in your email — that, choosing names, and approving what gets published are the only things it will ever need from you.
Two ways of working — you choose, and can switch anytime
Guide mode
It explains, then checks in
Before each step, the agent says what it's about to do and why, and waits for your OK — then runs it and reports back. Good for a first session, or whenever you want to follow along closely.
Agent mode
It runs, you approve what matters
The agent moves through routine steps on its own and narrates. It still stops and asks before anything consequential — publishing, deleting, spending.
If you run low on usage
Nothing is lost. Your work is plain files in your folder, and anything already published lives safely in your Looky instance.
- Wait for your usage to reset, or move to a plan with more room if you're working regularly.
- Come back later. The workbench keeps a short progress note, so the next session picks up where you left off instead of starting over.
Daily life in the workbench
Creating content is just conversation. "Make me a bar chart of sales by month", "a dashboard for the fulfillment team with these three KPIs" — the agent authors it, validates it, and shows you. For the mechanical operations, type @ in the chat and pick one:
- Set up the workbench — first-time setup, or repairing anything missing
- Connect to your looky — sign in / reconnect to your instance
- Choose a workspace — see and switch where you're working
- Download a workspace — bring a workspace's content to your folder
- Start a new workspace — begin a brand-new workspace
- Check for errors — validate everything before publishing
- Publish changes — review and publish, always with your explicit approval
- Switch working mode — between guide mode and agent mode
- Update this workbench — get the workbench itself up to date
What keeps you safe
- Nothing goes live without your yes. Publishing always ends in an explicit question, and the instance validates content before writing anything.
- Credentials stay out of reach. The workbench never touches data-source credentials or deploys source settings — that remains with your account's developer or owner.
- Everything is contained. The tools install inside the folder; your computer's setup is left alone. Deleting the folder deletes everything local (your published content lives in your instance — pull it again anytime).
See the change with your own eyes: after each publish, open your Looky in the browser and refresh. That loop — ask, publish, look — is the whole workflow.