Permission levels: Reads, Drafts, Acts explained

Every bot in our directory carries one of three permission levels. They answer the question you should ask before adding anything: what is the worst this bot can do to my accounts without me?

The three levels

• Reads. Reads your data and reports back. Never changes anything: no sending, posting, writing rows or creating drafts in your apps. Output goes to the chat or a file.
• Drafts. Prepares drafts and proposals. You approve before anything is sent or changed.
• Acts. Can send, post, move, pay or change things on your behalf. Read its risk flags.
• Not stated. The public description doesn't say. Check it in preview before adding.

The levels map to a pattern xAI's own guides use. In Grok Bot for Work, the author's Inbox Manager “could respond to everything,” but “I've set mine to read-only.” His Calendar EA drafts moves and invites “only after I confirm.” Same product, different levels, chosen on purpose.

How we assign a level

For listings imported from the official Marketplace or submitted to us, a rule-based classifier reads the public description, skills and integrations for action verbs (send, post, publish, delete, pay) and ignores negated phrases like “never sends without you.” That result is an estimate, shown with “est.” next to the level. A verified creator can state the real level, and a Hub test confirms it from what we saw the bot do; either one removes the “est.”.

What Grok Bot actually enforces

Our labels describe a bot. Grok Bot's controls decide what happens on your account:

A few defaults work in your favour:

• Messages are drafted by default. When a Bot prepares an email or Slack message, it shows an editable draft marked Needs approval unless you choose Disable drafts for Bot (docs).
• Email from a routine always waits as a draft until you press Send.
• Text from an email or a web page doesn't count as your request. Auto Review takes into account only what you asked for in your own messages.

The Grok Bot 101 guide is honest about the limits: if you log into Amazon on a computer the agent can use, “it can technically buy whatever it wants, the same way a human could.” Rules are written in natural language, and “you are trusting the model to follow what you wrote.”

Reading a risk flag

Flags name the specific capability behind a level: Sends messages, Posts publicly, Spends money, Deletes data, Uses your logins and Health data. A Drafts bot with “Sends messages” usually means it posts a summary somewhere you chose. An Acts bot with “Spends money” means exactly that.

Moving a bot up a level, safely

A level can change

Creators can update a template after publishing; the Share menu has Update template for exactly that. We re-check every listing's share link and log what changes on the bot's page, but the bot you already added is your own copy, so review its skills and routines yourself after any change you care about. To set up the guardrails, read Auto Review and approval cards.