Grok Bot: What It Costs, What It Can Access, and How to Build an Agent Team With It

Grok Bot is an agent product for handing work off across apps, websites and files, rather than asking questions in a chat. You give a Bot a role, a task and the access it needs; it works on a cloud computer with a browser, files and a terminal, and it keeps going in the background. SpaceXAI launched it in beta on August 11, 2026, on top of the Grok 4.6 model.

This guide covers that official product, not third-party Telegram or Discord bots using the Grok name. It answers the three practical questions first, because they decide whether the rest is worth your time: what it costs, what it can reach, and where it has to stop for you. Then it walks through the orchestration workflow a SpaceXAI engineer demonstrated publicly, which is still the clearest available picture of how agent teams get structured inside a frontier lab.

Grok Bot vs Grok: What Is the Difference?

Regular Grok is the assistant on grok.com and the mobile apps. Its documented capabilities cover conversation, research and writing help, file analysis, voice, image and video creation, and connected tools. It is not limited to simple questions.

Grok Bot adds a persistent, named AI agent organized around a job. Bots retain context, can coordinate with other Bots, and work in an agentic environment that includes a browser, files and a terminal. The product’s computer environment is powered by technology from Cursor, the coding startup SpaceXAI finished acquiring three days after launch.

The useful distinction is the handoff. “Help me draft a campaign report” is a conversational request. “Check the source dashboards each Monday, prepare the report, and leave it for my review” is a continuing workflow that needs access, a schedule and a defined deliverable. If your task is the first kind, start with the simpler interaction rather than granting extra access by default. Our guide to using Grok covers that side.

Start with one clear role rather than a set of vaguely defined helpers. SpaceXAI’s Bot-management guidance recommends separating roles when responsibilities differ, keeping stable instructions in the Bot’s description, and not letting Bot memory stand in for checking current source data.

What Can Grok Bot Do?

SpaceXAI’s own materials describe account research, outreach drafting, inbox organization, website work and customer support. Those are intended use cases, not evidence that every workflow succeeds unattended. For a first evaluation, pick a task whose result you can inspect.

Those are suggested test scopes, not results from hands-on testing. We have not independently benchmarked Grok Bot.

Bots reach software two ways: through connectors, or by driving the browser. SpaceXAI recommends a connector wherever one exists, because clicking through a website is less reliable, and reserves browser interaction for services without a connector. Either way, authentication can still need you: for passwords, two-factor codes, CAPTCHAs and payments, the documented pattern is to take control of the computer yourself, complete only the blocked step, and hand it back.

Working with X

An August 29 update added an X integration for searching posts, reading timelines, checking mentions and summarizing activity, along with a plugin for trends and bookmarks. That makes X research one available workflow. It does not make Grok Bot a route to unrestricted social automation.

Grok Bot Pricing and Eligible Plans

There is no standalone Grok Bot subscription. Access arrives with a qualifying plan, and Cursor’s live pricing page currently lists it on every paid individual tier.

These are subscription prices rather than quotes for unlimited work, and tax is additional. Cursor Teams and Enterprise are billed separately, and enterprise access is arranged through an account team. Linking a SuperGrok or X Premium+ subscription is a usage grant rather than a Cursor plan.

A documentation discrepancy worth knowing

The two official sources disagree about the entry point. SpaceXAI’s FAQ states that “eligible plans include SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams Standard and Premium,” omitting base SuperGrok and base Cursor Pro. Cursor’s live pricing page lists Grok Bot access on Cursor Pro at $20.

The table above follows the live pricing page, on the reasoning that a current commerce page reflects what is actually sold. Confirm the entitlement on your own account before paying for a tier specifically to get Bots, because the cheaper door may close as quietly as it opened.

Always-on does not mean unlimited

Paid access carries a weekly usage allowance. Work beyond it bills on demand where that setting is enabled, and combining a Cursor plan with a linked SuperGrok or X subscription does not stack the allowances. A job already running can finish past the monthly on-demand limit, so the cap is not a hard stop mid-task. The trial is a usage credit with a seven-day window rather than seven days of unlimited work.

Before an overnight job, check the remaining allowance and the on-demand setting together. A one-line instruction can start a very long task, and always-on agents consume capacity around the clock.

How to Use Grok Bot for the First Time

1. Install the official app. Setup covers macOS, Windows and Linux, with mobile available too.
2. Sign in with the account you intend to own this. Settle account ownership before linking plans or work tools.
3. Create one focused Bot. One name, one main responsibility, and boundaries written into its description.
4. Assign a bounded task. Specify the result, the sources, the restrictions, the deliverable and the review point.
5. Connect only the tools that task needs. Complete passwords, verification codes and other sensitive steps yourself through computer takeover.
6. Review the output before scheduling repeats. Fix missing evidence and unstated assumptions first.

A marketing team’s first brief might read:

Compare last week’s campaign performance with the previous four weeks using the dashboard I provide. Return the three largest changes, links to the supporting charts, and any missing data. Draft possible explanations separately from verified findings. Do not edit campaigns, change budgets, or send messages. Stop after the report for my review.

That is an illustrative brief, not a tested product result. Its value is that “done” can be checked.

The Access Boundary to Understand Before Connecting Work Accounts

This is the part most write-ups skip, and it is the one with consequences.

Every Bot on your account uses the same computer. SpaceXAI’s documentation is explicit that browser cookies and signed-in sessions are shared, files are visible to every Bot, and command-line credentials are shared. Creating a “Client A” Bot and a “Client B” Bot is an organizational choice, not proof that their data is isolated. Do not put material in that environment if another Bot on the account must not reach it, and apply the same caution you would to any shared workspace, as our piece on using AI tools at work sets out.

The approvals guidance recommends narrow rules for consequential actions: sending messages or invitations, publishing, purchases and transfers, deleting or overwriting data, changing permissions, production changes, and accepting legal terms. Auto Review is model-based and, in SpaceXAI’s own wording, should complement rather than replace least privilege and explicit approval boundaries.

Two more details belong in any security review. Grok Bot requires data storage and does not support Legacy Privacy Mode, with privacy and training choices following the applicable Cursor account settings. And deleting a Bot does not remove shared-computer files or browser sessions, so when access is no longer needed it has to be revoked in the connected service itself.

Skills and Routines: Repeat the Work Only After It Works

A skill describes how to perform a task. A routine assigns a workflow to a Bot and says when to run it. SpaceXAI’s guidance is to complete a successful one-time task before automating it, which is the right order for a reason: automation multiplies whatever the first run did, including the mistakes.

When you schedule a routine, define the time zone, the current input source, the expected output, the approval point, and what should happen when data is missing. A routine that quietly reuses last week’s numbers is not a report.

One warning deserves repeating: a test run performs real work. It can interact with websites, modify files and use connected tools. Treat a test as execution, not a harmless preview.

The Workshop That Started the Conversation

The most detailed public look at Grok Bot in practice came from SpaceXAI engineer Lauren Tan, who ran a nearly hour-long workshop showing how she works through 10 to 20 Bots, coordinated by a single Chief of Staff agent. Her claim is that the setup automates about 90 percent of her routine. Clips pulled in millions of views on X.

Tan carries real practitioner weight: she works on Grok Bot and Cursor at SpaceXAI, serves on the React compiler core team, and previously built software at Cursor, Meta and Netflix. Still, these were company demos during launch week, so treat the productivity numbers as a best case. The architecture is what makes the material valuable, because it is concrete and reproducible. (Individuals named in this article are referenced for news context only; they are not affiliated with Coursiv and do not endorse it.)

The Grok Bot Workflow, Step by Step

Step 1: Orchestration through one Chief of Staff

She does not message Bots individually for most work. She messages one, a Chief of Staff agent, and it delegates. The orchestrator knows which specialists exist, what each owns, and what is in flight, so it can route a vague request the way a human assistant would. The orchestration layer is the part worth copying, because it turns a pile of Bots into a team.

Step 2: Specialists with permanent roles

Under the Chief sit named specialists: an Inbox Manager, a Calendar Scheduler, a To-Do Organizer, and more as needed. The principle is persistent personas rather than one-off tasks. Each Bot owns a domain, keeps long-term memory of it, and improves at that domain precisely because the same thread keeps returning to it. Specialists can also message each other without routing everything through the orchestrator.

Step 3: Connect the ecosystem once

Her Bots plug into Gmail, Google Drive, Calendar, Notion, Slack, Salesforce, HubSpot and GitHub, across multiple accounts. Connections are account-level, so a tool connected once becomes available to every Bot, current and future. That convenience is the same mechanism as the shared-computer boundary above: connect deliberately, because the whole roster inherits it.

Step 4: Teach skills, not prompts

Instead of writing elaborate instructions, she teaches. Grok Bot’s Teach a Task feature records a workflow while she performs it once, then converts the demonstration into an executable skill. Skills also capture reusable preferences, down to details like scheduling household errands late at night, and they can be shared between agents. Combined with persistent memory, corrections stick: a Bot told once how she likes something handled applies it from then on.

Step 5: Routines, triggers, and the self-fixing loop

Saved skills become routines that fire on a schedule, like a morning brief that extracts what matters from newsletters, or on triggers, like an inbound message matching a pattern. The most striking example is the feedback loop Tan built because she wanted agents fixing bugs while she slept: when her team files feedback on Grok Bot itself, her Bots pick the reports up and start fixing the issues before she has read the ticket.

Step 6: Guardrails before autonomy

Every Bot in her setup runs with granular permission controls, and the recurring example is email: Bots draft everything, and nothing leaves for an external recipient without approval. Guardrails are what make an always-on agent usable. A Bot that must ask about everything is useless, and a Bot that never asks is dangerous, so the line gets written once, per Bot, in advance. Her own tip for controlling usage is scheduling discipline: a routine set to run every 15 minutes fires almost 100 times a day, while hourly or a few times daily is usually enough.

The Engineering Version: A Multi-Agent Workflow for Code

The same pattern extends to software work. A coordinating agent the team calls Tech Lead Taylor manages role-based engineering agents: a backend specialist, a frontend specialist and a QA persona. The demos show backend and frontend agents negotiating an API contract between themselves, flagging blockers and redistributing work as dependencies shift, while the tech lead agent tracks progress.

Output still flows through human review. Tan described merging around 20 AI-created pull requests in a batch, keeping the pull request process specifically because review and documentation are for people. She published her skill pack, Pstack, on the Cursor marketplace, reportedly shipped more than 1,000 pull requests in a month with it, and says she reviews outcomes rather than reading the code. An on-call agent watching Datadog and PagerDuty rounds out the roster.

SpaceXAI’s launch materials describe the same shape elsewhere: sales Bots for outbound research, lead scoring, demo setup and CRM hygiene; operations Bots for onboarding and invoice processing. The pattern is role-shaped, not job-shaped. Any function with recurring, multi-tool work can get a named specialist.

The Numbers, With a Grain of Salt

The claims are aggressive: 90 percent of a routine automated, and co-founder Jimmy Ba separately said 90 percent of SpaceXAI engineers work through agent loops and ship about five times faster. These are the company’s own numbers, on the company’s own workflows, demonstrated by experts during launch week. Early outside users report a bumpier picture: weekly allowances that heavy multi-Bot setups exhaust in days, costs that scale with always-on consumption, and fair questions about what routine actually needs 20 Bots.

The architecture is real and copyable. The percentages are marketing until independent users reproduce them.

How to Build Your Own Agent Team at Any Scale

You do not need 20 Bots. The pattern scales down cleanly.

1. Create one general-purpose Bot and treat it as your future Chief of Staff. Give it small, verifiable errands first.
2. Add your first specialist for whichever domain eats the most of your week, usually the inbox. Write its charter: what it owns, what good looks like, what it never does without asking.
3. Connect only the tools that specialist needs. You can widen the ecosystem later; you cannot easily narrow the blast radius of a shared computer.
4. Teach one recurring, multi-tool task by demonstration and turn it into one routine with a schedule or trigger.
5. Set the approval line by reversibility: Bots finish anything undoable on their own, and anything external, financial or permanent waits for you.
6. Review weekly. Ask each Bot what it ran, what it produced and what it skipped, then prune routines you would not miss.

Is Grok Bot Worth Trying?

It is worth evaluating when the work spans several tools, repeats over time, and produces a result you can verify. It is a poor first choice when you need a quick answer or a single draft, because that trade grants access you did not need.

A sensible first test measures three things: whether the output is correct, how much review it needs, and how much of the allowance the job consumed. Widen the scope when those three justify it, not because always-on sounds like set and forget.

The Skill Underneath the Workflow

Strip away the product and this is a lesson in management: defining roles, briefing owners, setting boundaries, reviewing output. That is the skill the whole agent era is converging on, and it transfers to every tool shipping this shape, from Grok Bot to Claude Cowork and whatever follows. It is also learnable before you spend anything on subscriptions. Coursiv teaches those foundations through short daily lessons on practical AI skills, built for people without a technical background. Check the official site for current course details and pricing.

FAQ

The Bottom Line

Grok Bot’s teaching is not that one engineer automated 90 percent of her routine. It is the shape of the system: one orchestrator, named specialists, skills taught by demonstration, triggers that make Bots present rather than punctual, and guardrails drawn at the point of irreversibility. Learn the shape at whatever scale you can afford, and before you connect a work account, be clear that the Bots you create are colleagues sharing one desk rather than tenants in separate rooms.