Security checklist for AI client work

The client's. Create the Supabase project, the hosting project, the repository, and every AI provider key in accounts the client owns, then have the client invite you. Anthropic's API key safety guide says: "If someone needs access to the Claude API, they should obtain their own key." The same page suggests separate keys for development, testing, and production, so a compromised key can be turned off for just that use.

Do not send keys to the client in a message. Supabase's API keys guide lists chat, email, and SMS among the ways a secret key must never travel. When the keys are created in the client's accounts, nobody has to send them.

If the work started in your accounts, plan the move before launch. The transfer steps for Supabase, Vercel, and GitHub are under How to fix.

Set the limit in the client's account at each provider. The controls differ, so here is what each provider's own page says.

OpenAI. The spend limits guide has two controls. A spend alert "Sends a notification; API traffic continues". A hard spend limit applies to the organization or to one project, and when it is reached, affected API requests "return a 429 error". To set one, go to Organization limits, or to the project's settings and then Limits. Select Edit spend limit, enter the Monthly spend limit, and turn on Enforce a hard limit. The guide says "Hard spend limits can interrupt production traffic" and "Enforcement is not instantaneous, so recorded spend can slightly exceed the configured amount." Add alerts below the limit so you hear about it first.

Anthropic. The rate limits page says the Start, Build, and Scale tiers each carry a monthly spend cap, and "You can also set your own spend limit below your tier's cap to control costs." Set it in the Claude Console on the Billing page, under Spend limits. When usage reaches a limit you set, "requests return HTTP 400 with error type invalid_request_error". If the client's organization runs more than one app, workspaces can each have a spend limit, set on the Workspaces page. That page also says "You cannot set spend limits on the Default Workspace," so give the app its own workspace.

Gemini API. Google's billing page says "You can set your own project-level spend caps in AI Studio." Set it on the Spend page, under Monthly spend cap, then Edit spend cap. The page marks project spend caps as experimental and says billing data can be delayed up to around 10 minutes, so spend can go past the cap before billing catches up. Long-running tasks such as batch mode completions and agent sessions may also exceed the project spend cap.

A provider limit is a backstop for the whole account. The app still needs its own per-user cap. See No cap on LLM API spend.

Write a short key note and hand it over with the app. One row per key: the service, the account that owns it, where the app reads it (for example, the host's environment variables), who rotates it, and how. Leave the key values out of the note.

For a Supabase secret key, the API keys guide gives the order. Create a new secret key in Settings, then API Keys. Replace the old key everywhere the app uses it. Confirm every component uses the new key. Then retire the old key. Deleting a secret key can't be undone, so do it last. The same guide suggests a separate secret key for each backend component, so a leak means rotating only that one.

For AI provider keys, Anthropic's API key safety guide suggests rotating "on a consistent schedule (for example, every 90 days) by creating new ones and deactivating old ones." Put the schedule in the key note, with the client's name next to it.