An agent edits your site. Here is every control you have on it.

Every CMS and integration credential you connect is encrypted with AES-256-GCM before it reaches our database, at every write path.

App servers on Hetzner (US). Database on Amazon RDS (US). Uploaded files on Amazon S3 (US).

Deleting your account schedules real deletion: within 30 days (after a grace period in which you can cancel), your workspaces, connections and stored credentials are permanently deleted. Billing records are kept only as tax law requires.

Agent traffic to your site comes from our server's published IP address, so your firewall can allow it and your logs can recognize it.

Your homepage, pricing, checkout, login, billing and legal pages are never edited automatically, in 7 languages. You can also protect any page, or a whole section with /landing/*.

Redirects, de-indexing, merges and deletions can never run unattended, in any mode. They also always require an admin.

Want image descriptions on auto but titles held back? Tick any capability and it always waits for your approval.

Unattended work stays under a daily credit cap you set. The off switch stops work in progress within one cycle.

Before any edit to an existing page, the agent saves the page as it was. Undo restores that exact version, from the task, from the activity feed, or with one workspace-wide button.

If you edited a page after the agent did, an automatic revert refuses to overwrite your work, and a manual undo warns you first.

A small automated change that measurably hurts its own page is rolled back automatically; bigger changes surface a suggested revert for your click. If a revert fails, a human on our team is alerted.

Every action lands in a permanent ledger: what changed, when, and the before and after. The receipt shows in your activity feed.

Workspace admins can export every agent change in a period: when and where it was decided (dashboard, email, chat, MCP, REST API or Slack), which account decided it when that is known, the quality review result, the AI models used, and whether it was undone. Optionally, name a Responsible Editor for your published text; the export includes it.

When the AI writes 150 or more new words, we keep a signed record of the text, and on WordPress and Shopify the page carries an invisible mark, as Article 50 of the EU AI Act asks. A public checker confirms a mark without revealing the workspace, site or page.

Moonshot (Kimi) is the main model, with DeepSeek as its outage backup, Anthropic Claude for specific jobs, and OpenAI and Replicate for images and embeddings. Optionally, a workspace can run its text generation on Anthropic or OpenAI using its own API key (the provider bills you directly; explicit recorded confirmation required). We use each provider's business API, never their consumer products.

Moonshot is China-hosted. Workspaces marked EU-sensitive are forced onto Anthropic in code - running on your own Anthropic API key, with an explicit recorded cost confirmation - so a routing mistake can never send their content to a non-compliant provider.

We never train models of our own on your data. Provider terms vary: Anthropic and OpenAI process API traffic under terms that exclude training on customer content; Moonshot's API terms carry no equivalent enterprise guarantee - which is exactly what the EU-sensitive toggle is for. Flag a workspace and its content runs only on Anthropic, on your own API key and agreement.

AI-written content must pass an independent quality review before it can go live; deterministic fixes carry a saved rollback instead. The agent refuses to invent facts, dates or links, and page content is fenced against prompt injection.

Cards only show figures computed from your own Search Console data, never a model's guess dressed up as a metric.