Product1 publisher3 min readPublished
Grok Bot ties write access to Gmail and Salesforce to a personal paid seat
Engadget says SpaceXAI's agent signs into Gmail, LinkedIn and Salesforce and acts for the user, with eligibility running through six individual paid tiers. The only brake described is a human-approval setting.
The Product Desk · Product desk
What happened
- Engadget reports that SpaceXAI opened the early public beta of Grok Bot in August, an agent that signs into tools like Gmail, LinkedIn and Salesforce and performs actions on the user's behalf.
- The bot fills out forms, responds on social media and sends email, then comes back with results or submits requests for approval.
- Access is gated behind six paid tiers, two on SpaceXAI and four on Cursor, so the beta reaches people through their own subscriptions.
- SpaceXAI says each bot learns by watching the user complete a task, retains that workflow as its own routine, and uses it to correct later runs.
- Grok Bot began as a Cursor project codenamed Sand and was partially built on SpaceXAI's compute. That history is part of why its name confuses people.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure A wrong run sends an email or alters a record, and the account that signed in owns it.
- constraint Because the entitlement follows an individual subscription, an employee can grant an agent write access to company systems without IT approving the connection.
- decision Approval is the only control in the product's own guidance, so teams have to decide which action classes require a human click and who is on the hook for clicking.
- contradiction The product is labelled a beta whose reliability may change, and the pitch is still unattended delegation. That leaves sensible use to work that is cheap to redo.
The overnight shift is the selling point. Engadget describes bots that message you, work with your tools, coordinate with each other and keep going while you are asleep [6]. A bot holding a live Gmail session and sending mail for you also means the first visible output of a bad run is an email in someone else's inbox with your name on it [2]. The naming does not help anyone explaining this internally: Engadget separates Grok Bot from the Grok chatbot, which answers prompts, and from Grok Build, the terminal agent that writes and ships code [11].
Eligibility attaches to the subscriber. The beta is open to SuperGrok Plus and SuperGrok Heavy subscribers, and to Cursor Pro+, Cursor Ultra, Cursor Teams Standard and Cursor Teams Premium [3], six paid tiers in total [12]. Two of the six carry "Teams" in the name and the other four do not [3]. Clients install on macOS, Windows, iOS or Android [4]. So the person who can connect an agent to the company CRM is whoever expensed a seat.
SpaceXAI says the bots learn by watching you complete a task, retaining that workflow as their own routine and using it to make corrections for improving future runs [7]. Whatever you demonstrate becomes the procedure. A shortcut taken while the bot watched, a skipped validation or a stale saved view, gets retained along with the useful part. Engadget's walkthrough covers plan eligibility and the human-approval setting. There is no admin console, no audit trail and no per-scope permissions in it [13].
Engadget's own advice is to begin with low-risk work and keep human approval enabled for anything that could fire off messages, alter records or expose sensitive information such as screen recordings [8]. That is a sensible default written for one user with one bot. As a policy for a team running several bots on the shared account computer, it leaves open who reviews the approvals [6].
Two questions sort the candidate tasks. Can the action be undone, and would a second person notice it happened. Undoable and noticed, such as a draft left in a shared folder, can run unattended. Undoable but unnoticed, such as a bulk field update in a CRM, needs a record of what changed before it gets autonomy. Not undoable and noticed, such as a LinkedIn post or a customer email, keeps approval switched on indefinitely. Not undoable and unnoticed is the box to keep a beta agent out of, and Engadget labels this one a beta whose features and reliability are subject to change [9]. Its suggested starting tasks, summarizing a set of documents or organizing research notes while you supervise [10], sit in the first box.
What to watch
- Whether SpaceXAI adds admin-side controls, per-scope permissions or an audit trail before the beta label comes off.
- Engadget flags evolving plan restrictions, so eligibility may move from individual subscriptions to org-level administration.
- Whether anyone other than the user who trained a bot can inspect and edit the routines it retained.