Grok Bot: Workflows and Limits

Grok Bot is worth trying for bounded research, reporting, and draft preparation, provided a person can review the result and recover a failed run. All your Bots share 1 cloud computer, and the subscription's weekly usage pool is shared with other eligible Grok and Cursor activity. Start with a job that can wait; do not make an unattended Bot the only owner of a business deadline.

Verified against SpaceXSI's live product, pricing, documentation, and changelog pages on 2 November 2026. This is a documentation-based evaluation, not a hands-on performance test. The workflows below are proposed operating procedures, not claimed test results.

What Grok Bot Actually Is

Grok Bot is an app from SpaceXSI, formerly xAI, for assigning work to persistent AI agents: software that can take actions across tools to complete a task. You create named Bots, give them jobs, and message them; they use connected apps or a cloud computer with a browser, filesystem, and terminal, the command interface for running software. Work can continue after your laptop closes, and the environment keeps files and signed-in sessions between tasks. The important distinction is that each Bot has a role and working context, while the computer belongs to your account and is shared across your Bots. Product overview.

Who Should Try It, and Who Should Skip It

Grok Bot fits an operator who repeatedly gathers information from several places, applies judgment, and prepares something another person can check. Think of a founder assembling customer-account briefs, an operations lead reconciling vendor updates, or a builder collecting evidence for a bug report. The work has a recognizable finish line, and a delayed or incomplete attempt is recoverable.

It is a poor first choice when the job requires guaranteed throughput, strict separation between client accounts, or unattended spending. A Bot that has the right answer but cannot finish before the usage reset is still the wrong system for a deadline. A named “Finance” Bot also does not create a protected finance compartment inside your account.

The decision turns on three criteria: can you inspect the output, can you restrict access at the source, and can the workflow survive a missed run? If the answer to any is no, keep the consequential step with a person or a separately controlled workflow.

Grok Bot vs Claude Cowork, ChatGPT Work, and Gemini Spark

Grok Bot's named roles are a useful organizing model, but background execution alone is not a reason to switch. These are the alternatives worth shortlisting:

Claude Cowork, Anthropic's task-capable workspace, is the first alternative for an existing Claude user. Cloud sessions and schedules can continue without a device online; work using local folders or the local browser has desktop dependencies. That makes “Grok works with the laptop closed” an incomplete comparison. See our Claude Cowork review and Anthropic's current surface guide.

ChatGPT Work, OpenAI's mode for carrying out multistep tasks, fits an operator who wants a finished result assembled from connected tools such as Gmail and Slack. Stay there if your existing plugins already cover the job; evaluate the review process and credit use before moving it. Our ChatGPT Work review covers the product; OpenAI's walkthrough explains when Work is appropriate.

Gemini Spark, Google's task workflow experience, is worth considering for personal-account workflows using connected apps, schedules, and browsing. Google's current help page presents these as Gemini tasks and still excludes work and school accounts. That is a substantial wall for a company rollout. See our Gemini Spark review for the account and workflow requirements.

Named Bots: Give a Recurring Job a Clear Owner

Grok Bot's persistent roles are most useful when the same standards matter every time. Create a Bot for account research and another for release monitoring only if they need distinct working instructions. A larger roster is not evidence of more work getting done.

In the sidebar, use + → Create a Bot, name it, and use the first message to specify its responsibility, sources, and stopping point. Lasting instructions can be added in conversation and remembered for later work. Create and manage Bots.

For a funded founder, a useful first role is Account Research. Hand it a deliberately short prospect list and a concrete instruction:

For these five named companies, prepare a brief with the product they sell, a recent relevant change, and the source links. Separate confirmed facts from guesses. Suggest a reason to contact each company, but do not contact anyone, add contacts to a campaign, or change the customer database.

The deliverable should make review easy: each statement has a source, each recommendation is separate from the evidence, and missing information is marked. Reject a fluent paragraph that disguises an unsupported assumption as a fact. The task succeeds when the founder can make an outreach decision faster, not when the Bot produces more text.

After reviewing the first brief, tell the Bot what should remain true on future assignments. For example: “Never call a company a fit without evidence of the purchasing need.” Keep task-specific details, such as this week's prospect list, separate from the durable rule.

Bots can coordinate and hand work to each other, but that adds another place for ambiguity to enter. Start with a single owner and add a second role only when the handoff has a defined artifact, such as a reviewed research file. A research Bot handing vague impressions to an outreach Bot merely moves the review problem downstream.

This capability's payoff is reusable context. Its wall is authority: a remembered instruction describes the job, but it does not limit what the connected account can technically access.

Computer Use and Connectors: Finish the Awkward Middle Steps

Grok Bot is a stronger candidate when a workflow crosses a supported connector and a website that still requires navigation. A connector is an authenticated app integration that gives the Bot specific actions; browser use lets it work through a site's visible interface. Those are different access paths, with different failure modes.

For an operations lead preparing a vendor review, the proposed workflow is straightforward: read the relevant email thread, inspect an authorized supplier portal, compare the current document with the saved previous version, and return a discrepancy list. The Bot should prepare the comparison, leaving contract changes and external replies for review.

Use Connect apps to add the required integration, finish authentication, and confirm it is installed and connected. An “Added” label alone does not prove sign-in succeeded. Attach the app with @ when giving the task. The Gmail connector can read and draft mail; it can also send, so a drafting assignment needs an explicit stopping point. Computer and app documentation.

A useful handoff names the exact email thread or folder, portal page, comparison document, and output destination. “Check everything with this vendor” leaves scope open. “Compare the attached renewal document with the approved agreement and flag changed obligations” gives you something you can verify.

Prefer a connector when its actions cover the job, then use browsing for the remaining steps. That is an operating recommendation, not a measured speed claim. Direct actions generally make the intended operation easier to describe and inspect; a browser path introduces page layout, session, and navigation dependencies.

The shared computer makes continuing another Bot's work convenient. It also means you should regard files and browser logins placed there as available to the whole roster. The diagram below shows the access model, not the product interface.

For a solo technical builder, keep the pilot in a staging account with disposable data. Have the Bot collect reproduction steps and evidence before asking it to make changes. If the job needs a privileged production login merely to inspect a problem, resolve that access design before handing over the workflow.

Skills and Routines: Automate a Finished Method

Grok Bot routines are useful after you have a process worth repeating. A skill describes how to perform the work; a routine starts it on a schedule or a supported event. Treat a successful first attempt as a draft method, then add the conditions that make later attempts safe to review.

A senior operator could start with a weekly account-health brief. The Bot gathers approved support and account signals, identifies changes, and leaves a linked watch list. It should flag missing source data, distinguish new evidence from an old note, and avoid contacting customers.

When Teach a task is available, you can demonstrate a browser workflow and review the resulting skill. Availability can roll out gradually; written instructions remain an alternative. A demonstration does not show what to do when a report is empty, a source is stale, or the login expires. Add those rules yourself. Skills and routines.

1. 1

   Run a small task once

   Use a restricted account list and demand source links. Check the result against the source systems. Correct the method before asking for a schedule.

2. 2

   Save the acceptance rules

   Ask the Bot to save the successful process as a skill. Include the output format, the definition of a meaningful change, and the rule that missing data produces an exception report rather than a guess.

3. 3

   Schedule and inspect

   Ask for the routine on weekdays at 8:30 in your office's time zone. Open the routine and inspect its owner, instructions, and saved schedule. A newly created schedule waits for its next scheduled time; use Test in the desktop app for an immediate run.

4. 4

   Review the run, then its consumption

   Confirm that a useful result arrived and inspect usage. Pause or narrow a routine that spends allowance without producing an accepted output. Do not solve an unclear task by running it more frequently.

Event triggers deserve the same restraint. A trigger for a particular escalation phrase in a particular channel has a clear purpose. A listener for every message may wake the Bot repeatedly for irrelevant conversation. Connecting an app for reading and configuring an event trigger are also separate setup steps in the docs.

Specify what happens after partial completion. For example: if a source is unavailable, return the available evidence, name the missing system, and stop. If a later retry would send a duplicate message or create a duplicate record, keep that action manual until you have a way to detect the previous attempt.

The benefit is less repeated coordination. The limit is that a saved schedule does not promise a completed result: source availability, usage, approvals, and run failures still matter.

Approvals and Vault: Access Is a Business Decision

Grok Bot's approval system can stop an action for your review, but you need to define which actions matter. Auto Review, the automated action check, is enabled by default. It can allow work to proceed without a human prompt; it is not a promise that every consequential action will present a confirmation.

For a paid-media operator, a reasonable task is to reconcile performance and prepare a proposed change. Ask for the current setting, proposed setting, evidence, and affected campaign. Require a human decision before the Bot changes a budget or sends the recommendation outside the company.

In Settings → General → Bot, keep Auto Review enabled and add narrowly described Ask first rules under Custom Rules. For example, require review before sending external messages or changing production settings. Turning Auto Review off also stops your personal custom rules from being used. The vendor describes the review as model-based, so permissions in the source service remain essential. Approvals documentation.

There is an operational catch for unattended work: approval cards raised by routines or other background work expire after 30 minutes. A request started directly in chat does not have that timer. Schedule approval-dependent work when someone can respond, or have the routine produce a proposal that a person later asks it to execute.

A morning briefing is an appropriate overnight job. A purchase that must complete before a supplier closes is not the same kind of delegation. The person who owns the deadline needs a fallback if the Bot stops halfway through.

Vault Protects Secret Values, Not Separation Between Bots

Grok Bot's Vault stores logins, cards, and contact information and fills secret values without showing the password or card number to the Bot. Use the secure secret card for API keys and the secure handoff for sign-in steps. Do not paste credentials into ordinary chat or leave them in working files. Vault entries are account-wide and cannot be restricted to a single personal Bot. Vault documentation.

Before connecting anything, ask: would it be acceptable for any Bot on this account to use this access? If a marketing research Bot must never touch payroll, the payroll login does not belong on that shared computer or in its account-wide Vault. Naming another Bot “Payroll” does not change that answer.

Use a restricted source account where the service supports it. Give a research workflow access to the relevant material, not the founder's unrestricted mailbox by default. Account permissions that prevent a change are a firmer boundary than an instruction asking a model to remember not to make it.

Offboarding is separate work. Pause related routines, sign out of websites, remove files, and revoke integrations at the source. Deleting a Bot alone does not remove the shared computer's browser sessions or files. Otherwise an experiment can leave working credentials behind after its owner believes it has ended.

Desktop and Mobile: Review Away From Your Desk

Grok Bot's desktop apps run on macOS, Windows, and Linux; its mobile apps run on iPhone and Android. They reach the same Bots, conversations, routines, connectors, and cloud computer. Mobile is useful for assigning a follow-up, answering a question, and reviewing an approval while away from your desk. Mobile documentation.

For an operator between meetings, the sensible mobile task is bounded: “Compare these supplier options and leave the tradeoffs here.” The phone is then a place to inspect a finished comparison. It should not turn every spare minute into approving poorly scoped actions whose consequences require a larger screen and source documents.

Mobile and desktop are not identical in every control. You can inspect and manage routines on the phone, but testing a routine requires the desktop app. Set up and exercise a new recurring workflow at your desk before relying on mobile notifications to keep it moving.

Grok Bot for Windows

Grok Bot for Windows is available through the official download route at x.ai/bot, alongside the other desktop installers. Follow the app's browser sign-in and use the account that holds your entitlement. The getting-started guide describes Cursor authentication and the separate access paths for linked Grok subscriptions.

The cloud computer is distinct from your laptop. Local command execution is an additional permission with its own controls. For a pilot that only needs cloud work, leave local execution disabled; add it only for a specific task that needs local files. A cloud agent does not need unrestricted access to your own machine to justify its subscription.

What Grok Bot Costs to Run

Grok Bot's sensible entry point for a new individual pilot is Premium, the lowest current X Subscription tier that includes it. The cheaper Plus plan does not cross that access boundary. These are the monthly web prices shown on SpaceXSI's pricing page on 2 November 2026, before applicable taxes:

• Free: Grok Bot is not included.
• Plus, $8/month: Grok Bot is not included.
• Premium, $30/month: Grok Bot is included; start here for a bounded individual pilot.
• Super, $100/month: Grok Bot is included with a larger shared allowance; upgrade only after measuring the job.
• Ultra, $200/month: Grok Bot is included with a larger allowance again; it is not a published guarantee of completed tasks.
• Teams, from $30/user/month: eligible paid Premium, Super, and Ultra seats include Grok Bot.
• Enterprise, custom pricing: access and usage controls depend on the agreement.

For the full subscription ladder, read Is Grok Free?. The relevant buying decision here is the work you can finish within your spending limit, not the number of features elsewhere in the bundle.

Grok Bot Cursor Access and the Weekly Meter

Grok Bot consumes the weekly pool allocated to your paid seat on The X Subscription, shared across eligible first-party Grok and Cursor surfaces. SpaceXSI does not publish the absolute size of that pool. It cannot honestly be translated into a promised number of tokens, messages, jobs, or hours. Different activities consume different amounts. Usage limits.

On self-serve Teams, each seat has its own weekly allowance; those pools are not combined. Paid Additional Usage draws from a shared prepaid team balance. For individuals, top-ups also fund activity beyond included usage. Check Auto Top-up and spending caps before leaving a routine unattended. Third-party models in Cursor are separately paid Additional Usage, outside the included first-party pool.

Existing subscribers should check entitlement before buying again. The docs list legacy paid Cursor plans, eligible linked SuperGrok and X Premium+ subscriptions, and self-serve Grok Business seats as access routes. New customers start on The X Subscription. Moving eligible old accounts into the unified subscription changes the allowance model, and the documentation says unification cannot be undone. Subscription access and migration.

Measure Accepted Work, Not Bot Activity

The useful cost calculation is subscription cost allocated to the workflow, plus top-ups and the cost of human review, divided by accepted outputs. An accepted output meets the standard you set before the run; a draft you must substantially redo does not count.

As an illustration, allocating the entire $30 Premium subscription to 20 accepted briefs in a month gives $1.50 per brief before top-ups and review labor. If only five briefs are usable, the same allocation is $6 per brief. These are arithmetic examples, not forecasts of how many briefs the allowance will support. If you use the bundle for other work, allocate its cost consistently rather than charging every workflow the full subscription.

A pilot log needs the job, accepted result, review time, extra spend, and usage readings before and after. Because other Grok and Cursor work shares the meter, a before-and-after reading during mixed activity is only an approximation of the Bot's consumption. A busy dashboard is not a business outcome.

Where Grok Bot Breaks

Grok Bot's practical limits concern capacity, access, and recovery. The adoption risk is not just a wrong answer: it is a partially completed job whose owner thinks the automation has it covered.

The November 1 Cost Signal Makes Capacity Harder to Trust

On 1 November 2026, Elon Musk said the cost of serving Grok Bot was “unsustainably high relative to revenue”, following complaints about unusually fast usage exhaustion. His statement did not announce a new price, specify an allowance, or promise a reset. Contemporary reporting described users finding their weekly meters exhausted or sharply increased overnight; those are user reports, not a controlled measurement of a night's work. Reported complaints and context.

The operational implication is uncertainty about how much work a subscription can support over time. Cost pressure may affect included limits or how usage is allocated, but the statement does not establish what changed for any individual account. It is not a basis for predicting a price increase or declaring every meter jump legitimate consumption.

For a founder, that means retaining the manual route to a critical report and observing a full weekly reset cycle before expanding the schedule. A subscription that worked for a burst of experimentation has not yet demonstrated dependable capacity for recurring commitments.

Separate Bot Names Do Not Isolate Client Data

A consultant working for unrelated clients cannot use separate Bot names as the separation mechanism. The personal account's computer and access remain shared. Resolve confidentiality requirements through permitted accounts, separate environments, and source permissions before loading client material. Do not discover the access boundary after the Bot has already read the wrong folder.

Blocking an App Does Not Close Its Website

The computer-and-apps docs make an easy-to-miss distinction: blocking a connector does not block browser access to the same service. The documented control for closing the website path is Enterprise Network Controls. A mid-market CTO evaluating “we disabled that integration” needs to check both paths. A policy that covers only the connector leaves a different route available.

That does not mean the Bot automatically bypasses policy. It means an app block and a network restriction enforce different things. Validate the actual boundary with the administrator before calling a service inaccessible.

Scheduling Does Not Supply Recovery

The routines documentation names failed runs that are not automatically retried after a rate-limit error. A recurring schedule is therefore not equivalent to a monitored production workflow. Someone must notice the missing output and decide whether retrying is safe. Put an owner and a fallback beside any result a customer or colleague expects.

Approvals can also interrupt work while nobody is watching. Browser sessions can expire, and sites can require human authentication. A workflow should report which steps finished and which did not, rather than returning a confident summary that conceals an incomplete operation.

Persistent Does Not Mean Every Installation Survives

The cloud computer retains working state, but the docs distinguish ordinary software updates from replacing or recovering the computer. Full computer updates, Recover, and Reset can remove installed apps and packages; recovery can interrupt active work. Keep important outputs in the destination system, and do not make an undocumented installation the only record of how a business process runs. Computer maintenance behavior.

For a technical builder, write down the dependencies before relying on them. For a nontechnical operator, make the Bot's output understandable enough that another person can continue the job without reconstructing its environment.

Verdict: Try a Bounded Role, Wait on Unattended Ownership

Grok Bot earns a pilot when a job is repetitive enough to reuse context, varied enough to benefit from judgment, and limited enough for a person to check. The strongest initial uses are research briefs, reconciliations, monitoring summaries, and drafts with explicit source links and stopping points.

Adopt it for a workflow only when accepted output saves more effort than setup, review, and recovery consume, and the observed usage leaves room for the rest of your seat's work. If reliability depends on an undisclosed allowance lasting until a particular deadline, or on named Bots having separate access, wait.

For next week, choose one recurring preparation task. Create its owner Bot, connect only the necessary sources, keep payment cards and privileged production logins out, and run a small example while present. Review it, then observe the same bounded task across a weekly reset. Expand only when you can explain both the useful result and what happens when the run stops.

A founder already satisfied with Claude Cowork or ChatGPT Work should test a specific missing workflow before switching. A personal Google-account user can consider Gemini Spark. An operator needing guaranteed throughput or strict client separation should keep that work in a system whose capacity and permissions can be explicitly controlled.

Frequently Asked Questions

What is Grok Bot?

Grok Bot is SpaceXSI's app for persistent named AI agents that work across connected tools and a shared cloud computer. Each Bot can have its own job and working context, while account-level files and browser sessions remain shared.

How much does a Grok Bot cost to use?

For a new individual subscriber, Premium is the lowest current X Subscription tier that includes it. The plan lines above give today's prices. Usage beyond the included weekly allowance can require paid top-ups, so the subscription price alone does not establish the cost of an unattended workflow.

How do I get a Grok Bot?

Download the app from the official Grok Bot page, sign in with your eligible account, and create a Bot with a specific job. Desktop and mobile access use the same account; existing legacy subscriptions may already grant access.

Is the Grok Bot actually good?

It is a credible candidate for preparation work that crosses apps and benefits from persistent context. This review verifies documented behavior, not task success rates. Decide with a bounded pilot that measures usable outputs, review effort, and consumption before trusting recurring work.

Is Grok bot free to use?

The current Free and Plus tiers do not include Grok Bot. Do not confuse a free download or free Grok chat access with an ongoing Bot entitlement. Check the official plans and your existing subscription before paying again.

Get the AI Tools Map for Business Owners to match the next workflow to the right tool.