2026-08-27 · Tutorial
How to Test Grok Bot on the Trial Without Wasting the Credit
Opening Grok Bot on a trial and clicking every connector is how you finish the sample before you have a single page you would hand to a colleague. The 21 August 2026 trial exists. It is limited usage, consumed by agent work and tokens. It is not a week of unrestricted teammates. Cursor billing writeups sometimes describe a seven-day window. Treat that as a rumor until you read the terms on the screen in front of you. This page will not invent a dollar figure for the credit, because none is published on the SAFE list.
This is not what the trial is, which covers eligibility and the meter. This is not how to test a bot setup, which is a golden-set method for a bot you already intend to keep. This page is how to spend the sample: one reversible job, no send, no pay, inspect the output, then stop. For the product from scratch, see what a Grok bot is.
Name the sample job before you open the Grok Bot app
Write one sentence a skeptical teammate could grade: this bot will produce X from Y public sources, and it will not do Z. If the sentence needs "and" twice, you have two jobs. The sample is sized for one.
The job has to be work you were going to do this week. "See whether Grok Bot is cool" has no pass condition. "Build a five-company lead sheet for Thursday's pipeline review" does: five rows, each with URLs you can open, in a document you own by morning.
Write the stop verb in the same sitting. Never send, purchase, create accounts, sign into mail or a bank, or post. An approval controls the next proposed action. It does not reverse work already completed. If you cannot undo it, do not spend the credit on it. Approval rules and reversibility is the longer axis.
Name the artifact: a table or a brief with a source list, not "insights." Fluency is free. A reusable page is the only thing the credit is for.
Pick the machine before the run. On iPhone (iOS 18+) you can pause and resume a routine, read its run history, delete it, and open the Agent Computer. Edits and testing need a desktop, on macOS, Windows or Linux. Android phones and iPads run the same companion limits. The agent runs on a managed Linux VM in the cloud, which is not a Linux desktop client.
Read the meter as limited usage, then confirm the window in the product
SpaceXAI widened access on 21 August 2026 and described a one-time trial for individuals. The useful fact is the shape: limited usage, not unlimited runs until a calendar date. Some Cursor billing pages talk about a seven-day window around that usage. Confirm the current terms in the product and on Cursor pricing the morning you start.
There is still no published numeric credit. Posts that print a dollar amount or a token count are guessing. Plan as if browsing is expensive. Loading pages, waiting, and retrying on a cloud desktop costs more than one API call for the same fact. Token burn with no per-Bot spend cap is the paid-plan version of that warning. On the trial, the same physics applies with a smaller tank.
If you already hold Cursor Pro, Pro+, Ultra, a Cursor Teams seat, or a linked SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription, this is not a trial. You already have the product. Cursor Hobby, the free plan, still does not include Grok Bot, and an unlinked SuperGrok does not count until you link it. Privacy Mode (Legacy) blocks Grok Bot entirely.
Leftover credit is a successful sample, not a reason to start a second experiment. The Gmail cookie you create tonight will still exist tomorrow.
Pick a reversible assignment that cannot send, pay, or publish
Reversible, for a trial, means two things. The output can be thrown away without anyone else seeing it. The side effects on the shared computer are ones you can live with if you abandon the product at lunch. Public-page research passes both. A Gmail login fails the second even if you never hit send.
All bots on an account share one persistent cloud computer assigned to the user, not to a bot. Each bot gets a screen. Screens are not security boundaries. Cookies, sessions, files, and CLI credentials are shared. Removing a bot does not wipe those. Hosted MCP sign-in tokens stay with Cursor's backend. Browser sessions stay on the computer. See shared computer security. Do not sign into anything you would hate leaving on that VM.
| Candidate sample | Send or pay? | Login required? | What you learn | Use on the trial? |
|---|---|---|---|---|
| Five-company public lead sheet with URLs | No | No | Whether the browser and citations work | Yes, this is the job |
| Inbox drafts that never send | No send if the charter holds | Yes, mail | How Gmail feels, plus a session on the VM | No |
| Slack standup posted to a channel | Yes, a post | Yes | Whether the room notices | No |
| A purchase, a signup, or a form fill | Yes | Often | That the bot can click | No |
Least privilege is the standing rule after you pay. On the trial it is narrower: public web, one bot, one artifact, then stop. Lead Scout is the catalog shape for research that never contacts anyone. Inbox Triage and Chief of Staff Briefing want mail or calendar access, so save them for a plan you will keep.
Run overnight lead research on five public companies with sources
You sell a product that ops or security teams buy. You already have five public companies on a maybe-list. You were going to spend an evening on them. The bot will, overnight, from public pages only. You inspect the packet in the morning and then you stop.
Use five names you actually care about. A worked public list: Atlassian, Salesforce, ServiceNow, Adobe, HubSpot. Replace any name with a company on your real list. Do not add a sixth. Five source URLs is a morning you can click through.
Write the columns before the bot runs, so you cannot grade fluency later:
| Column | What belongs there | What does not |
|---|---|---|
| Company | The legal or traded name you typed | A nickname the bot invented |
| Public trigger | One fact that would change a sales conversation, with a URL | A claim with no link |
| Org or hiring signal | One careers or IR page, with a URL | A guess about headcount |
| Fit note | Two sentences labeled as your hypothesis | "Great fit" with no reason |
| Next human step | What you will do | An email the bot drafted to their CEO |
Evening: twenty minutes to paste the charter and name the five companies. First five minutes of the run: watch the Agent Computer for IR pages and newsrooms, not a LinkedIn login. Then close the lid. Overnight: one run. Morning: thirty minutes to open every URL. Then the job is over, even if usage remains. A second job "because there is credit" is how people connect Slack at 8:12 a.m.
If you cannot leave the overnight run, you did not test Grok Bot. You tested a chatbot. Public IR, newsroom, careers, filings, blog only. If a source is blocked, write blocked and move on. No account creation, no mail, no "workaround" login.
Leave Gmail, Calendar, and Slack disconnected on day one
The other way to spend the sample is the one launch videos teach. Connect Gmail, Calendar, and Slack. Ask the bot to "just see my day." You will learn that plugins have a browser auth flow. You will leave three sessions on a shared computer. You will have no artifact except a feeling.
Grok Bot and Gmail and scheduling assume you intend to keep the computer. If the agent cannot research five public companies, mail will not save it. If it can, mail is a later grant on a plan you chose with a price in front of you.
| Day-one stack | What it costs in the sample | What it leaves behind | What it proves |
|---|---|---|---|
| Public research, one bot, no plugins | One bounded browse | A file you can copy off the VM | Whether citations and the computer work |
| Gmail plus Calendar plus Slack | Auth, inbox crawl, calendar crawl, chat crawl | Three sessions, shared across every future bot | That connectors open |
| Ten bots in a group chat | Coordination overhead on a small tank | A credential spread you did not mean | The demo, not the job |
A routine assigns a workflow to one bot, with a maximum of 50 per bot and 20 recent run records kept. Removing the bot removes its routines. Clocking a five-minute job all night converts a sample into an empty meter. Run once. A Slack login on the trial VM is a Slack login for the next bot you create if you upgrade on impulse. Do not use separate bots as a security boundary.
Paste a trial charter that fails loudly when the job is done
The charter is the whole test. If it is vague, the run will browse until the meter is sad. Paste this. Change only the five company names and the one-line product description. Do not add "also check my mail if you have time."
Name: Trial Lead Sheet
Job: Five public companies, one sheet, then stop
Product I sell: meeting-prep software for ops teams (replace this line).
Companies, exactly these five, no extras:
1. Atlassian
2. Salesforce
3. ServiceNow
4. Adobe
5. HubSpot
For each company, fill one row with:
- public trigger (one fact, one URL you opened)
- org or hiring signal (one fact, one URL you opened)
- fit note labeled HYPOTHESIS, two sentences
- next human step that I will do, never you
Use public IR, newsroom, careers, blog, and SEC filings only.
If a page is blocked or wants a login, write blocked and skip that cell.
Do not create accounts. Do not fill forms. Do not use LinkedIn while signed in.
Do not open mail, calendar, or chat.
Boundary: never send email, never post, never purchase, never sign into
sites, never use cookies from other bots, never continue browsing after
the five rows exist.
Deliverable: one markdown table, five rows, then halt.
If you finish with time left, do not start a sixth company.
Say the boundary out loud before you paste it. If you cannot say "it never sends" to a person who does not care about agents, the charter is not ready. "It is careful" is not a boundary. Repeat the stop list in the first message. Save the brief outside the app so an upgrade starts from the same five-company job, not from a new wish list.
Inspect the morning packet, then stop even if credit remains
Morning is the test. Not the overnight log. Not the feeling that it was working when you went to bed. Open the packet. Count the rows. You asked for five. If you have three, the run failed completeness. If you have twelve, the run failed the stop line. Both are useful failures. Neither is a reason to connect Slack.
Open every URL. If a link 404s, or points at a homepage with no mention of the claimed fact, that cell is a fail. Fluent prose over a dead link is the failure mode this job is designed to catch. You are not grading writing. You are grading whether the computer retrieved a page and told the truth about it.
Look at the Agent Computer at least once. You are hunting logins you did not authorize and form fills you did not ask for. The phone can open it too, but a desktop screen is where you read it properly. Use desktop. Then stop. Leftover credit is evidence that you named a job small enough to measure. Is Grok Bot worth it is the verdict after this packet, not before it. Do not upgrade to avoid wasting remaining usage. There is still no Grok Bot-specific spend cap. Copy a good packet into a document you own. The trial computer is not an archive.
Score citations and coverage, not fluency
A generous reader will call any tidy table a success. Do not be generous. Score the sheet the way you would score an intern's first research dump: every material claim has a URL, the URL supports the claim, and the row is complete enough to decide a next human step.
| Check | Pass | Fail |
|---|---|---|
| Row count | Exactly five | Fewer, or extras "for context" |
| Trigger URL | Opens, and the fact is on that page | Homepage, 404, or a fact not on the page |
| Second URL | A different page | Duplicate links, or "source: knowledge" |
| Blocked cells | Marked blocked | A confident paragraph where the bot was blocked |
| Outreach | None | A drafted email, a form fill, a follow on X |
| Stop | Halted after the table | "I also checked your inbox" |
You need four of five trigger URLs to work, and zero outreach, for a product pass. Three working URLs and two blocked cells is a maybe. Zero working URLs and a beautiful narrative is a fail. Homepage summaries are what a chat window already does. If the Agent Computer spent the night on four homepages, fix the charter or walk away. Do not "fix it" by connecting CRM. Date the score next to the charter. Cursor Pro at $20 a month is the cheapest documented individual paid door. An individual SuperGrok subscription can be linked instead. A self-serve Cursor Teams seat includes Grok Bot for every member. None of those prices is a reason to upgrade if the sheet failed.
Keep plugins, routines, and extra bots off the trial VM
Each plugin starts with an Add click and a browser login. On a trial, that login is a credential on a VM you might not keep. The safety checklist is written for people who are staying. Act like you are leaving.
One bot. Extra bots get extra screens on the same machine, not extra computers, which means extra ways to inherit a cookie. Group chat is how a research bot and a mail bot share a session. The trial is one worker on one task.
Teaching by showing you click around records up to ten minutes of the screen, no mic, browser flows only, a draft skill at the end, and it does not run on iPhone. Writing the charter is cheaper. Do not drop CLI tokens onto the trial computer. Those credentials are shared. If a plugin is required for the job you named, you named the wrong trial job.
Answer the claim that a product tour is the only honest test
The strongest objection to this method is honest. You are not buying a research chatbot. You are buying a teammate that can sit in Gmail and Calendar and Slack. If you never connect those, the trial is a lie. You tested a browser, not the product.
That objection wins if your only question is whether connectors open. It loses if your question is whether you should pay. Paying is a weekly allowance plus on-demand overflow, on a shared computer, with no published numeric cap and no audit view of bot actions outside Enterprise. Send, post, purchase, and public schedule are the parts you must not spend a sample on, because an approval does not undo them.
Once Gmail is signed in, every later bot on the account can inherit that session. You no longer tested research. You tested research on a machine that can also read mail. The trial is too small to afford a contaminated computer. Tours also have no pass condition. A five-row sheet does. You can fail it.
If you already know you will buy Cursor Pro+ because your company standardized on it, you can tour after you pay. The trial is the door for people who should not pay yet. If your inbox is not allowed on a vendor-managed desktop, skip rather than connecting mail "just to look." Feeling the product is the first five minutes of watching the Agent Computer. Finishing the sheet is the test.
Trace a burned sample back to the click that spent it
When the meter is empty and the packet is thin, the cause is almost never "the model." Grok Bot has no model picker for members or admins. You can pin a job. You cannot pin a version.
| Symptom | Likely click | Repair, if any credit remains | Repair, if the sample is gone |
|---|---|---|---|
| Credit gone, no table | Unattended crawl with no stop line | Stop. Copy any partial file. | Browsing is expensive. That is a result. |
| Beautiful prose, dead links | No URL contract | Demand the table in one more short run | Do not upgrade on prose |
| "It sent a message" | Missing send boundary, or a plugin | Kill the connection. Treat as a failed test. | Revoke at the provider. |
| Gmail still signed in after you removed the bot | You connected mail. Removal does not clean sessions. | Revoke in Google. | Same revoke. Removal was never the cleanup. |
| iPhone-only trial you cannot inspect | You started on mobile | Finish on desktop | You tested a pause button. |
| Routine firing every five minutes | You scheduled the sample | Delete the routine. | The usage is already gone. |
If anything left the building, the trial is over as a test and open as an incident. Check sent mail, Slack, calendar invites, and form confirmations, then revoke. A blocked IR page is sometimes the VM's datacenter address. Mark it blocked. Do not solve it by logging in to look more like a human.
Prove the trial with checks a generous reader cannot fudge
Write the pass bar before the overnight run, then grade against it in the morning. If you invent the bar after you see the prose, you will pass yourself.
The artifact exists in a document you own, not only inside the product. You opened at least four of five trigger URLs and the claimed fact was on the page. The bot did not send, post, purchase, or sign in. You watched the Agent Computer at least once. You did not add Gmail, Calendar, or Slack. You did not create a second bot. You copied the charter out. If you connected anything anyway, you have already opened the provider's third-party access page and revoked it.
That list can fail. A trial that cannot fail those checks taught you nothing you could defend in a meeting. Grok Bot cost is what you read if the sheet passed and you are considering a plan. Read the vendor pages the same day.
Search your vault for logins you minted during the sample, including a throwaway inbox you opened "just in case." Close it. Assume the computer stored secrets until you have looked. If you cannot spare the morning inspection, do not start the overnight run.
Name the cases where one overnight job is the wrong sample
This method has a domain. Outside it, a different test is honest, or no test is.
If you already hold an eligible paid plan, you are spending weekly allowance, not a one-time sample. You can still use the five-company job as week-one discipline.
If the only work you would ever give a bot is sending mail, use the research analog, or skip. Do not "just send one test email to myself." That is a send. If you need a golden-set and injection tests, that is testing your bot, after you pay.
If you only have a Linux desktop, use the Linux app. If you only have an iPhone, an Android phone, or an iPad, you can pause, resume, read run history, and open the computer, but you cannot edit or test. Wait for a desktop, or skip.
If your five companies are private, pick five public analogs and label the sheet as analog research. Do not sign into a data vendor on the trial VM.
If you needed Claude Code skill compatibility, you are on the wrong surface. That is Grok Build, not Grok Bot. The Grok Bot docs do not describe reading SKILL.md or CLAUDE.md. See Grok Bot versus Grok Build. If you have not answered Cursor account questions (whose workspace, whose payment method), answer those on paper first.
Export the method to a note you keep after the window closes
Copy four things into a note you own: the charter, the five company names, the morning scores, and the revoke checklist. That note is the onboarding doc if you upgrade, and the record that you did not leave a bank session on a VM if you do not.
Revoke at the provider, not only in the bot list. Open the connected-apps page at Google and Slack even if you think you stayed on public pages. Watch the computer, then revoke, then copy the sheet off the VM.
Write the upgrade rule before the meter hits zero. Upgrade only if you used the artifact in a real meeting or review, and you have revoked every unused login. Supported platforms will not expand because you paid. The shared computer will still be shared. The missing per-bot spend cap will still be missing.
If the sheet passed and you would run this weekly, you have a candidate for a paid bot with the same send boundary. If it failed, stay on the chat product you already have. Compare Grok Bot versus Claude Cowork or Grok Bot versus ChatGPT Work only after you have a packet. The last line of the note is the stop line you already used: never send, never pay, inspect, then stop.
Keep reading: What the Grok Bot trial is covers eligibility and the meter, how to test a bot setup is the golden-set method after you keep the product, and token burn with no per-Bot spend cap is what you live with if you upgrade.
Frequently Asked Questions
Can I treat the Grok Bot trial as unlimited use for seven days?
No. The August 21 2026 path is a usage sample. Agent work and tokens spend it. Some Cursor billing writeups describe a seven-day window around that usage. Confirm the current terms in the product rather than on a screenshot. A long unattended browse can empty the sample before morning. Plan one reversible job, inspect the packet, and stop even if the meter is not at zero. Exact credit amounts are not published on the SAFE list, so ignore posts that print a dollar figure they cannot source from the vendor.
What is the best first job when I test Grok Bot on the trial?
One overnight lead sheet on five public companies, with a URL on every material claim, and a charter that forbids send, pay, signup, and mail. That job measures the cloud computer and the browser without leaving a session you will regret. Replace the five names with companies you already needed to research this week. Do not add plugins. Do not schedule a routine. Do not start a second bot. If the morning table has working links and no outreach, you tested the product. If it does not, you still learned not to pay yet.
Should I connect Gmail, Calendar, and Slack while I test Grok Bot on the trial?
No. That stack is a product tour. It burns usage on consent screens and inbox crawls, and it leaves sessions on the one computer shared by every bot on the account. Removing the bot does not remove those sessions. An approval does not reverse a send that already happened. Connect mail after you are on a plan you intend to keep, and after a public-web job has already produced an artifact you reused. The trial is for a reversible sample, not for living inside your real tools for a day.
How is spending the trial different from a full bot test plan?
Spending the trial asks whether one reversible job produces a packet you would reuse, before you pay. A full test plan, the kind in the testing-your-bot guide, assumes you already intend to keep the setup. It uses a golden set, adversarial cases, and a boundary the bot must refuse. That work is right after you upgrade. It is the wrong way to empty a metered sample. Use the trial to buy a measurement. Use the later plan to buy trust. Mixing them is how the credit disappears on tests you will just rerun on a paid week.