Agent-safe publishing: preview, then approve.
An AI agent can do the work of putting a site online. It shouldn't also get to decide when the public sees it. Thora splits the two.
Why a preview and an approval.
Agents are fast, and sometimes wrong. They can misread a request, break a page, or act on text they found in a file or a web page that they shouldn't trust.
A private preview lets the agent make its mistakes where no customer can see them. One human approval means a person looked at the result before the public did. Together they let you hand over the work without handing over the last word.
The path to live.
Work on a branch.
Your agent's changes go to a separate branch. Nothing it does there touches the live site.
Test in private.
Thora builds a preview with its own database. Where it's switched on for your organization, your agent opens it signed in, tests it and reads the build logs. Only your team can open it.
Request the release.
Your agent asks for the exact build it tested to go live. It can ask, but it can't approve its own request.
You approve.
A person reviews the request in the Thora dashboard and approves it once. One approval also applies that release's tested database changes, in order, and then visitors get the new version.
What an agent can and can't do.
An agent can
- Create a site and change its files on a branch
- Build a preview, test it and read build logs
- Set up the site's database and run its changes in the preview first
- Ask for a secret by name, and get a secure link for you to use
- Save the live hostname, and request a release or a rollback
An agent can't
- Approve a release or a rollback, even its own
- See the value of a secret you enter
- Buy anything or change billing
- Go beyond your role or the organization you chose
- Gain permission from text in a file, a page or a tool result
Secrets go through secure links.
Never paste a password or a key into a chat with an agent. A chat can be stored, shared and read back later.
When the site needs a secret, your agent asks Thora for it by name and gives you a link. You sign in again, type the value on that page, and save it. The agent never sees it. The site reads the secret when it runs, so it never sits in the code.
A way back.
Thora keeps earlier releases. If a release is wrong, your agent can request a rollback to an earlier one, and a person approves it the same way. A request is not the rollback: it happens when you approve.
Rolling back changes which version visitors see. Don't count on it to undo changes to your data. If a release changed the database, ask your agent what recovery is available before you approve.
What to look at before you approve.
The request in your dashboard is short on purpose. Check these.
- That the site and the hostname are the ones you expect
- That the build is the one your agent tested in the preview
- Any database changes the release will apply
- That your agent told you in plain words what changed
Honest limits.
Thora Cloud is in beta. A real person sets up each new workspace by email, so there is no self-serve checkout yet.
Approval catches what you can see. It doesn't replace a look at the preview. Open the preview yourself before you say yes.
Some tools depend on your plan and on what's switched on for your site. Your agent checks what's available and says so when something isn't.
Keep reading.
Claude is a trademark of Anthropic. ChatGPT is a trademark of OpenAI. Thora Cloud is independent and isn't affiliated with or endorsed by either company.