AI agents, MCP and permissions

What agents can do

The MCP connection lets a compatible assistant read this website's content and numbers and, when the plan permits, prepare drafts. It cannot publish, send or schedule email, charge money, refund, delete or change settings. Those remain explicit dashboard actions for an authorized person.

Read-only access is included on Starter; Growth and Pro include read plus draft under the current standard plans. Effective account overrides may differ. Check AI agents for the current entitlement, token limit and request allowance rather than treating this article as a billing quote.

Create a token

Open Settings → AI agents as an Owner or Admin. Name the token after the client, choose Read or the available Read + draft scope and an expiry. Copy it when created: the secret is shown once. Keep it in the client's protected credential configuration, not in a public repository, screenshot or conversation.

Use the smallest scope and shortest useful lifetime. A website token is tied to its website; it is not an account-wide password. Verify which website you selected before connecting a client.

Connect all websites in one account

Only the billing-account Owner can open Account settings → AI agents and create an All websites in this account token, after recent authentication and two-factor enrollment. It covers current and future websites in exactly that billing account, never another account even if you own both. Use Read by default; Read + draft still depends on the current plan. Account tokens have a separate active-token cap from website tokens.

Configure the same MCP endpoint and Bearer header with this token. First call list_websites, choose the intended website, then include its website_id explicitly in every website tool call. There is no remembered current website. Both account and website tokens share the selected website's request allowance and the account safety budget. Revoke the account token in Account settings when it is no longer needed.

Configure the assistant

Follow the AI assistant walkthrough. The dashboard provides copy-paste configurations for Claude Code, Claude Desktop, Cursor, Hermes Agent and OpenClaw. Other clients need MCP Streamable HTTP, the displayed server address and an Authorization Bearer header containing the token.

Restart or reconnect the client, then begin with a read request. Review attributed agent drafts in the dashboard before saving or publishing anything. A prompt asking an agent to perform a forbidden action does not grant that permission.

Limits, downgrade and revocation

All tokens for one website share its tool-call allowance, and an additional account safety limit applies. Adding tokens does not multiply the rate. If rate limited, wait for the returned retry time.

A downgrade does not delete tokens or drafts; an existing draft token can still read but cannot draft if the plan no longer allows it. Revoke unused or exposed tokens on the same screen. Revocation blocks future requests; it cannot erase information already copied by a client. Create a replacement rather than trying to reveal an old secret.