How to ship an app with a team of Grok Bots

A writer built and shipped a working app called Quiet Desk in three days using a team of Grok Bots — a PM Bot, Design Bot and QA Bot — plus Cursor coding agents, sending the PM Bot 42 messages, about half of which were just "yes," "OK" or "done." Grok Bot, launched in August 2026 by SpaceXAI on Mac, Windows, Linux and phones, lets users create named AI agents that work on a cloud computer and hand coding jobs to Cursor agents, which SpaceX acquired that same month. The app, which turns GitHub Markdown repos into a readable book-like view with a map showing 248 related file pairs in a repo with no internal links, went live on the third evening, with the first build starting 20 minutes after the idea was sent; the cheapest plan including Grok Bot is Premium on The X Subscription at $30 a month, and the whole run fit inside it with free hosting.

Over three days, I ran a small product team from a chat window. A PM Bot wrote the brief and ran the project. A Design Bot chose the look. A QA Bot tested every version and reported what broke. Cursor’s coding agents wrote all the code. I sent the PM Bot 42 messages. Half of them just said yes, OK or done. Let me show you the result first. It started with a habit. Lately, I’ve moved more and more of my work into Markdown files. They’re spread across many GitHub repos. And GitHub is a poor place to read them. You click through folders, one file at a time, and each file stands alone. My files aren’t always well organized either. The knowledge was all there, but hard to explore. So I imagined a simple app. You connect one of your GitHub repos. Then you read its Markdown files like the pages of a book. A file tree on the left takes you to any file in one click. On the right, the app suggests related files, even when they don’t link to each other. That’s how you find links between notes you never connected yourself. A map view shows how all the files connect. And you keep private notes on the side, without changing anything in GitHub. My Design Bot named it Quiet Desk. Here it is, live, on a public repo of design skills. Switch from Read to Map, and the repo turns into a map of ideas. Each dot is a Markdown file. A solid line joins two files that link to each other. A dotted line joins two files the app finds related. This repo has no links between its files at all, yet the map shows 248 related pairs. Those dotted lines are the connections I’d never have found by browsing GitHub. Click a dot to open its page, or type a name to highlight it. It went live on the third evening. I spent a few hours on it myself, in short sessions between other things. The Bots and the coding agents worked far longer, often at night while I slept. From sending my idea to the first build starting, it took 20 minutes. So what is Grok Bot https://x.ai/bot ? It’s an app from SpaceXAI, the team behind the Grok models. It lets you work with a team of AI agents. It launched in August 2026, on Mac, Windows, Linux and phones. That same month, SpaceX bought Cursor https://thenextweb.com/news/spacex-cursor-acquisition-completed-gpu-fleet , the AI code editor. That’s why you sign in with a Cursor account. You create Bots and give each one a name and a job. Then you chat with them like colleagues. They work on a computer in the cloud, so they keep going with your laptop closed. When a Bot needs code, it hands the job to one of Cursor’s coding agents. Using it feels like a WhatsApp group with a few fast, chatty colleagues. They reply quickly, even in the middle of a task. Their messages are short and to the point, sometimes too short. Once, PM Bot asked whether my repo was public or private. I replied “I don’t understand your question.” It explained in plain words, and we moved on. Some actions were surprisingly quick. Design Bot sent a full design direction a minute after it was asked. Others were slow, like the first build, which took over two hours. And watching the Bots message each other to move your project forward is fun. What does it cost? The cheapest plan with Grok Bot is Premium on The X Subscription https://x.ai/news/x-subscription , at $30 a month. The $8 Plus plan includes Cursor https://x.ai/pricing , but not Grok Bot. If you already pay for an older Cursor plan, you keep access. One weekly pool of usage covers Cursor, Grok Bot and the Grok apps. Every Bot reply draws on it. My whole run fit inside Premium, and the hosting was free. This post has four parts. - The seven steps. Here’s how I set up three Bots, and how they took my idea live. - The verdict. What the run cost me in time, tokens and money, and what a review of the code found. - Beyond prototypes. My 20-plus Bots for everyday product work, and what other PMs hand to theirs. - Under the hood. The questions I still had after weeks of use, from the model to what’s safe to connect. By the end, you’ll know how to take your own idea live with a team of Bots. You’ll also know what to watch along the way. You’ll have an app you can open on your phone, built while you did other things. The code Cursor’s agents wrote for my Bots is public at The-AI-Thinker/quiet-desk https://github.com/The-AI-Thinker/quiet-desk . Every source I used is listed at the end, under Go further . Three hires, and no interviews. Let’s set up the desk. Seven steps from idea to live app The first three steps build your team. The last four ship the app. Step 1: Set up Grok Bot You set up Grok Bot once, and every new idea starts from the same app. It’s a desktop app, and a few of its settings live on Cursor’s website. Ten minutes here saves you usage later. You set it up in this order. 1. Download Grok Bot from x.ai/bot https://x.ai/bot . It’s the app where you create your Bots and chat with them. Sign in with your Cursor account, since there’s no separate Grok Bot login. 2. Check that your plan includes Grok Bot. Premium does, and so do the older paid Cursor plans. Plus doesn’t. 3. Open Cursor’s dashboard https://cursor.com/dashboard , then Cloud Agents , then Default Model . Cloud Agents are Cursor’s coding agents, which your Bots start for each build. They use the model you pick here. 4. Keep Grok 4.7. A model from another provider bills at its API list price https://cursor.com/docs/models-and-pricing from the first request. 5. In the same menu, switch Fast off. It’s on by default and costs twice as much https://cursor.com/docs/models/grok-4-7 per token. 6. Connect other apps only when a Bot asks for one. Every Bot can use every app you connect. How I did it. I started on my old Cursor plan, Pro+, and switched to Premium during the run. Once it was over, I moved down to Plus. My coder ran on Grok 4.7 High, with Fast off. Speed wasn’t the bottleneck in my run. I was. The desk is set. Now hire the one Bot you’ll talk to. Step 2: Create your PM Bot One Bot owns the outcome, and it’s the only Bot you talk to. SpaceXAI’s own docs say to give one Bot ownership of an end-to-end outcome https://docs.x.ai/grok-bot/bots . Your PM Bot asks the questions, writes the brief, briefs the others and reports back. A Bot is an AI teammate with a name and a job. You give it that job in the chat, then ask it to remember. Creating it takes a minute. 1. Click New , then Create new Bot , and name it PM Bot . 2. Send it a role like the one below, in a few lines, adapted to your idea. 3. End with Remember this role , so it saves the role to its memory. 4. If you want a hard stop on risky actions, open Settings , then General , then Auto-review . Add Ask first rules for merging, deploying and publishing. That last step matters because a rule in a Bot’s memory only guides it. Cursor’s team, on its forum, calls Ask first the only setting that actually pauses an action https://forum.cursor.com/t/175615 . How I did it. I wrote the three role prompts before opening Grok Bot, and sent them as written. Here’s the PM’s. From now on, you are my PM Bot. Remember this role. Your job: take one product idea from me to a live app. You own the outcome and you’re the one I talk to. When something is unclear, ask me, one question at a time. You never write code. For any build or fix, you start a Cursor Cloud Agent and send me its screenshots, video and pull request. You work with Design Bot and QA Bot. Message them only when you need them, and keep updates short. Ask for my yes before you merge, deploy or publish anything. It saved the role, then warned me it couldn’t reach a Design Bot or QA Bot. They didn’t exist yet. Five minutes later, once I’d created them, it could. From then on, PM Bot asked before every merge and every deploy. A Bot can’t check its own work. Time for two more. Step 3: Add a designer and a tester The coder will tell you its work is done. Someone who never writes code has to check that claim. That’s the QA Bot’s whole job. The Design Bot sets the look before the build, then reviews each version. SpaceXAI’s docs say to add a Bot only when the work has a stable specialist role https://docs.x.ai/grok-bot/bots . These two qualify. All your Bots share one computer in the cloud. You get one per account, and every Bot uses its files, its browser and its logins. Here’s how the four of them work together, with you on top. You create both the same way as the PM Bot. 1. Create a Design Bot and a QA Bot , each with a short role. 2. Tell the QA Bot it never changes code or settings. Its job is to use each version like a first-time user, then try to break it. 3. Tell both to reply only when someone @-mentions them. Every Bot reply counts https://cursor.com/help/grok-bot/group-chats toward your usage, replies between Bots included. 4. Skip the group chat if you like. Your PM Bot can message the other two directly. How I did it. Here’s the designer’s role, sent as written. From now on, you are my Design Bot. Remember this role. Your job: make the app a pleasure to read and explore. Before the build: propose one design direction layout, typography, colors, how moving between files feels , with 2 or 3 reference apps known for a great reading experience. After each build: review the screenshots and the preview, and list the 3 changes that would improve it most, most important first. Reply only when someone @-mentions you. And here’s the tester’s. From now on, you are my QA Bot. Remember this role. Your job: use each new version of the app like a first-time user, in your browser. Follow the main path end to end, then try to break it. Report each problem with a screenshot, the steps to reproduce it, and how serious it is. Never change code or settings. Reply only when someone @-mentions you. I didn’t use a group chat. PM Bot messaged the other two itself. A small “Messaged Design Bot” line showed up in our chat each time. The team is hired. Now give it something to build. Step 4: Turn your idea into a brief The brief is what the coder builds and what the tester checks. Your PM Bot writes it from a short interview. Your job is to answer, then read every word it wrote. Here’s how the interview goes. 1. Send your idea in a few sentences, and list what the brief must cover. Mine asked for five things: who it’s for, the one job of the first version, what’s in and out, what “done” looks like, and how to check it. 2. Answer the questions one at a time. Yours will differ from mine, since each session is a little unpredictable. 3. Read the brief before you say yes, because the coder builds every word of it. Watch for words the PM Bot added to your answers. 4. Say yes when the PM Bot offers to get a design direction from the Design Bot. How I did it. I’d prepared the idea in advance too, and sent it like this. Here’s the idea. Turn it into a one-page brief for v1, as a Page. Ask me what you need to know first. The idea: I have a lot of .md files in my GitHub repos, and they’re hard to read and explore there. I want a web app where I connect a GitHub repo and explore its .md files in a beautiful, flexible way: read them like pages, see how ideas connect across files, and take notes. No Notion: the files stay in GitHub. The brief says who it’s for, the one job of v1, what’s in and what’s out, what “done” looks like, and how we’ll check it works. PM Bot asked six questions in 15 minutes, one at a time. Was it just for me? Should notes stay in the app? Should the app find related files on its own? A minute after my third answer, the brief was a page in Notion. PM Bot sent the brief to Design Bot. The direction came back a minute later, and it had a name: Quiet Desk. It came with three reference apps, a serif font for reading and a warm paper background. It read like a spec from a designer with opinions. The brief and the look were approved. Next, the code. Step 5: Let your PM Bot start the build Grok Bot doesn’t write code itself. Once you approve the brief and the look, your PM Bot offers to start the build. It writes the coder’s instructions and starts a Cursor Cloud Agent. The agent builds on its own machine, and your PM Bot brings back what it made. Your part in this step is small. 1. Say yes when the PM Bot offers to start the build. It writes the coder’s instructions from the brief and the design, so you don’t have to. 2. Let it run, because a first build can take a couple of hours. The PM Bot’s card links to the agent’s page, if you want to watch it work. 3. Read the PM Bot’s report when it lands. Expect screenshots and a list of what isn’t finished. Often there’s a video of the app at work too. 4. Decide where the code lives. By default it stays on Origin, Cursor’s own code host, which is fine for a test. To start on GitHub instead, connect it in Cursor’s dashboard https://cursor.com/dashboard under Integrations . You can also ask a Bot to publish the code on GitHub later, as I did. How I did it. After the design came back, PM Bot asked whether to start the build. I said “Yes please continue” , and it started the agent itself. The agent worked for 2 hours and 15 minutes. Then PM Bot brought back a walkthrough video, five screenshots, and a list of what wasn’t finished yet. The video was a nice surprise: I watched my app work before I’d even opened it. Videos kept coming after each round of fixes, so every change was easy to check. That’s a working first version, from an agent I never talked to. Step 6: Let the Bots test and fix The coder says the app works. Your QA Bot checks that claim against the brief, and the fixes loop until it agrees. Anthropic calls this pattern evaluator-optimizer https://www.anthropic.com/research/building-effective-agents : one model builds, and another evaluates and sends feedback. In my run, the Bots ran this loop among themselves. My part was mostly saying yes. Here’s the loop. Here’s what you do while it turns. 1. Say yes when the PM Bot offers to bring in the QA Bot. It sends QA the build and the brief’s checks. 2. Read the PM Bot’s summary of each QA report. It turns the findings into one fix list. Then it asks before it sends the list to the coder. 3. If the small questions pile up, tell the PM Bot to decide those itself. Keep your yes for merges, and for anything that touches your accounts. 4. The QA Bot may ask for something, like a test repo on your GitHub. Ask what it’s for, then ask it to clean up afterwards. 5. Merge when QA says ready. On Origin, a Bot can open a pull request but can’t merge it. So you click Merge yourself. A pull request is a proposed change you approve before it joins the main code. How I did it. PM Bot sent the first build to QA Bot. Four checks passed. Then came two serious bugs. Once anyone opened a private repo, any visitor could read the saved copy. And one huge repo froze the whole server. The coder had called this version done. PM Bot turned QA’s findings into a brief for the coder. All I typed was “Yes please send them” . Its brief was better than mine would have been. QA Bot’s retest came back in capitals: “NOT READY TO MERGE” . The two security fixes held. But two fixes the agent claimed hadn’t landed. Six pull requests went through this loop. I gave each round a few minutes, often at night, between other things. At one point, I asked PM Bot to stop asking me small questions and work on its own. It still asked before every merge, and I clicked Merge six times. Once, QA cleared a pull request at 3:48 in the morning. My yes came at 8:14. The team was faster than its manager. QA said ready. Time to put it online. Step 7: Put it live Hosting means picking a place for your app to run, at a cost you accept. Your PM Bot can lay out the options. It can also do most of the setup in its own browser. You keep two things: the logins and the final yes. Here’s how it goes. 1. Say yes when the PM Bot offers a hosting plan, and ask for a free option. Free plans exist, with limits like a server that sleeps. 2. Let it do the setup in its own browser. When it needs a login, it hands you its screen. A bar then says You’re in control . Sign in, then click I’m done . The Bot doesn’t see what you type. 3. When the Bot asks for something you don’t understand, ask why. It’s a conversation, so push back when a plan doesn’t fit. 4. Say yes to the deploy. Then let the PM Bot check the live app end to end. It signs in, opens your data, changes something, restarts, and confirms the change is still there. 5. Open the live app yourself, on your laptop and on your phone. 6. If Chrome calls the site Dangerous , ask the Bot to request a review from Google. Until it clears, warn anyone you send the link to. How I did it. PM Bot’s plan recommended Render https://render.com at about $25 a month. So I asked, “How about a free hosting solution?” It found one: Render’s free plan, with my notes in a free Neon https://neon.com database. Free has a price, though. After 15 quiet minutes, the app takes about a minute to wake up. And repos stop at 400 Markdown files. Then I typed “Do the setup in your browser, I’ll help with logins” . Neon went through on its own. Render needed me, because my account refused a GitHub login. PM Bot handed me its screen. Then it wanted to create two apps on my GitHub account. Before saying yes, I asked, “Why do you want to create a developer app on my account?” The answer was plain. GitHub only lets a site see your files through an app you register. I don’t accept an AI’s request until I understand the why. Same as with the hosting price: ask, challenge, then decide. At 10:24 pm, a card asked Deploy Quiet Desk to Render now? I said yes. The deploy took under two minutes. PM Bot’s own check then wrote a test note, redeployed, and found the note still there. Then two things broke. First, signing in sent my browser to a server-only address, 0.0.0.0:10000 . QA Bot had warned about it hours before, in a list of minor notes nobody acted on. One more fix round sorted it. Second, Chrome showed a full red page: Dangerous site . That comes from Google Safe Browsing, which Chrome uses by default https://support.google.com/chrome/answer/99020 . It had flagged my site as phishing, a fake login page that steals passwords. Mine stole nothing. But a brand-new site with a GitHub button and a broken redirect looked the part. PM Bot asked Google for a review, which can take a few days to a few weeks https://support.google.com/webmasters/answer/9044101 . It’s live, and QA says ready. Time for the verdict. The verdict on Quiet Desk Here’s my quick review of the run: my time, the money, the tokens and the code. Then what I’d change. My time came to about three hours. That’s my rough count from the message timestamps, over three days. The Bots and the agents did the long hours, often at night. The money: my plan, plus 11 cents. The run used 338 million tokens, almost all included in Premium. Is that a lot? At Cursor’s rate https://cursor.com/docs/models/grok-4-7 for usage beyond a plan, the coders’ Grok 4.7 tokens alone cost about $90. In plan terms, it’s a big chunk. By Oct 23, my weekly meter read 57%, and that week began after the first build. My rough read: a project this size takes most of a Premium week. So run one at a time. Where did the tokens go? More than nine in ten were the Bots and agents re-reading the conversation so far, before each reply. What they actually wrote came to under 1%. A long thread costs more than a long answer. So keep Bots quiet unless someone @-mentions them, and start each new idea with a fresh team. The code is a solid first version for one user. I asked Claude Code to review the public repo against my brief. That’s another AI, so take it as a second opinion. Here’s its summary. A senior engineer would accept it as a working v1 for one user. They would not accept it as a base to grow on until CI exists and the top three problems are fixed. It praised the security work. The app checks your access before it shows a page, and it never writes to GitHub. It also found three problems QA Bot missed. Here they are, in plain words. - One odd character can block a whole repo. If any file holds a link with a % sign, like 100%-done , the repo won’t open at all. One bad file should fail alone. - The automatic checks only pass on one setup. The agents wrote 103 tests, small programs that check the app still works. They pass on a machine like QA Bot’s, but 16 don’t pass on my Mac. And nothing runs them on each change, which developers call CI. So a future fix could quietly break something. - “Never touch my GitHub” isn’t locked in the code. For private repos, the app asks GitHub for a key without asking for read-only. The key gets whatever access was ticked in GitHub’s settings. Today the code only reads. One wrong tick, and a future version could write. What I’d do differently. Quiet Desk was a quick experiment to test Grok Bot, not a product. Still, next time I’d change three things. 1. Try the main feature on my own files before saying yes. The agent tested related files on the Rust book, never on my notes. It links files that share words, so two notes on one idea, in different words, stay apart. 2. Put my real setup in the brief. My repos live in GitHub organizations, and the app only shows my personal repos. So I could only test it on a scratch repo. 3. Ask for the outside review before the first merge, not after. Then have the tests run on every change. What it would take to share it. If I wanted other people to use Quiet Desk, the list gets longer. - Sign-in for more than one person. Today it accepts one GitHub account: mine. - Repos in GitHub organizations, where most real notes live. - The three fixes above, with tests that run on every change. - Paid hosting, so the app never sleeps and opens big repos. PM Bot’s first plan was about $25 a month. - A clean bill from Chrome, since nobody trusts a link flagged as dangerous. My take: for three hours of my time, it’s good enough to use and to show. Turning it into something others rely on is a real project, with a sharper brief. That’s the app. But I’d used Grok Bot long before Quiet Desk, for my everyday work. Use Grok Bot beyond prototypes A few weeks before Quiet Desk, I started a week with my usual rule. Use the new tool for everything, even when it wastes time. So each time I’d normally open Claude Code or a chat, I created a Bot instead. Within days, I felt it could replace my own AI setup. I ended up with more than 20 Bots, one per task. - One read the daily meeting transcript and commented on the tickets it touched. - One turned user interview transcripts into research reports. - One drafted my Slack messages. - One translated technical terms into plain words. - One turned the questions I get about my product into a Q&A. - One drew diagrams. - One reviewed Figma mockups and wrote a UX audit. Each one understood its task from a few lines of context. I just chatted with it, and it got the job right. I got a little addicted. Product people who run Grok Bot every day have learned things I hadn’t. Here are five I’d copy. 1. Give each Bot one job, and a name to match. Kevin Niparko works on product at SpaceXAI. His Bots include a chief of staff, an engineering manager that never codes, five engineers, a data analyst and a PM. He prefers specialists https://x.ai/bot/guides/grok-bot-for-pms to one Bot that does everything. You know who does what, they work in parallel, and each learns its own job. Claire Vo names each Bot after its job. Her pull-request Bot https://www.youtube.com/watch?v=QBmgF1kJSK4 cleared about 50 waiting pull requests in a day, by her count. 2. Teach it your voice before it writes for you. Grok Bot writes short by default, as I found. Niparko’s best instruction: “Read my past Slack messages and develop a Slack style guide for me” . Claire Vo asked hers to learn her voice from her emails. Niparko still sends most messages himself, and edits most docs. 3. Let a Bot track where your time goes. Niparko’s chief of staff reads his email, Slack, meeting notes and calendar through the day. It keeps a short list of what he’s actually working on, updated hourly. Then he asks it where his time went, and which of that work an agent could take over. 4. Turn the team’s work into updates. A Cursor engineer, Amrita, has a Bot that turns pull requests into release notes https://www.youtube.com/watch?v=rkigdXf-52I for customers. After a launch, Grok Bot’s product team asks a Bot for an hourly pulse on installs https://www.youtube.com/watch?v=x5OawRpQxII . 5. Don’t become your Bots’ full-time manager. Brian Casel, a solo founder, warns against checking your Bots’ work in chat all day. His Bots email him https://www.youtube.com/watch?v=RM8ZwhoJdeg their reports and approval requests, and he answers them in one reply. Niparko asks his to stay quiet unless something needs him. Even after weeks of this, I still had basic questions about how it all works. Look under the hood Which model answers me? Where are my Bot’s instructions? What happens when I hit send? Is it safe to connect my accounts? I asked myself all four, after weeks of daily use. Try Grok Bot yourself, since that’s the best way to learn it. You’ll ask them too. Here are the answers I found in SpaceXAI’s docs, on Cursor’s forum and from people who use it most. Which model runs your Bots You can’t pick it. Grok Bot uses the model “most likely to deliver the best outcome” for each task, says Cursor’s help page https://cursor.com/help/grok-bot/models . That can be an xAI model or another provider’s, like Opus 5.5. It can change with every request. There’s no model picker, and no effort setting for a Bot’s own replies. Grok Bot used to run on Grok alone. Since it moved to a mix of providers, two users on Cursor’s forum say it got “dumber” https://forum.cursor.com/t/174784 . Your only model control sits on the coding side. That’s the Cloud Agents default model from Step 1. Even that control is partial. After the run, I exported my usage from Grok’s dashboard. The file names the model behind every coding request. A fifth of my coders’ tokens went to Claude 4.5 Sonnet, which I never picked. Cursor’s staff explained it on their forum https://forum.cursor.com/t/169638 . Some helper tasks still pick their own model, “like exploration, browser or computer-use, and verification” . My first coder tested the app in a browser and recorded a video. That’s the work they mean. On the Grok Bot side, the file leaves the model blank on every row. The model is hidden from you. So are the instructions. Where a Bot’s instructions live When I created my Bots, I looked for a page with their instructions. There isn’t one. Your role prompt goes into the Bot’s memory. You can’t open that memory to read it. AIMultiple’s comparison https://aimultiple.com/always-on-agents found no documented way to view or edit it. Cursor’s team adds that only “the newest few thousand characters” of it reach each reply forum https://forum.cursor.com/t/175615 . A long, growing list of rules pushes the oldest ones out of view. The memory is kept apart from the chat, and survives a chat reset, say Cursor’s staff https://forum.cursor.com/t/168333 . Working files sit on the Bot computer. SpaceXAI’s docs suggest a shared folder, https://docs.x.ai/grok-bot/computer-and-apps /workspace , with one clear folder per project. What surprised me is how little a Bot needed. Each role was about five lines long, and each Bot stuck to its job for three days. So keep each role short, and ask the Bot to repeat it. Try What do you remember about your role? Longer rules can sit in a file on the Bot’s computer. The Bot reads it only when it opens it. Memory is one piece. Here’s what happens when you hit send. What happens when you send a request When you message a Bot, the work doesn’t happen on your laptop. Cursor picks a model for that request, and the Bot answers. Then it works on its cloud computer. It browses, runs commands in a terminal, or uses an app you connected. It can also message another Bot, which wakes, handles the request and replies later https://docs.x.ai/grok-bot/chat-and-collaboration . For code, it starts a Cloud Agent on a separate machine. Who coordinates it all? No built-in manager does it behind the scenes. The coordinator is a Bot you create and give that job, like my PM Bot. And nobody checks its work for you. When another Bot sends it a result, “nothing checks that result against your rules” , say Cursor’s staff forum https://forum.cursor.com/t/175615 . Catching a mistake is your job, or a QA Bot’s. That’s why mine earned its place. All of that work happens on one machine, and every Bot you create shares it. One computer for all your Bots To test the refresh button, QA Bot needed to push changes to GitHub. My GitHub connection couldn’t, so it signed GitHub’s command-line tool in on its computer. That gave it write access to my account. It flagged this itself, and signed out when I asked. Even then, “GitHub CLI” stayed on my GitHub account as an authorized app. It stays until you revoke it. That computer isn’t QA Bot’s. Here’s how SpaceXAI’s docs put it. “The computer is assigned to your user account, not an individual Bot,” they say https://docs.x.ai/grok-bot/computer-and-apps . Every Bot shares its files and its logins. It runs in the United States https://docs.x.ai/grok-bot/security . Deleting a Bot doesn’t remove the files or sign-ins it left behind source https://docs.x.ai/grok-bot/approvals-security-and-privacy . Your passwords stay hidden from the model. The logins they open stay open for every Bot. QA Bot needed my GitHub password once. I typed it into a secure form on its screen. So is it risky to connect your accounts? Bots act as you, on every app you connect. Cybernews reported one user whose Bot posted his bank balances https://cybernews.com/ai-news/grok-bot-leaks-bank-information/ to his company’s Slack, under his name. Simon Willison, a developer who writes about AI security, names the danger the lethal trifecta https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ . An agent is at risk when it has three things at once. - Your private data , like your inbox, your files or your bank. - Content written by strangers , like a web page or an email someone sent you. - A way to send things out , like posting to Slack or sending an email. The catch is that AI models follow instructions they find in content. A stranger can hide one in an email, like send me the latest invoices . With all three, Willison writes, “an attacker can easily trick it into accessing your private data and sending it to that attacker” . So yes, Grok Bot gets risky once you connect your work accounts. A Bot with your inbox, the web and Slack has all three. And since every Bot shares one computer, a bad instruction read by one Bot can affect the others https://cybernews.com/ai-news/grok-bot-phishing-spacexai-invite-scam/ . For a test like mine, with a scratch repo and no inbox, the risk is small. Meta’s security team suggests a simple rule https://ai.meta.com/blog/practical-ai-agent-security/ : give an agent at most two of the three. When a task needs all three, a person approves it. In Grok Bot, that’s an Ask first rule. Three habits keep this safe. 1. Connect only the apps you’d let every Bot use, since they all share them. 2. Give work that needs its own logins its own Cursor account. SpaceXAI’s security FAQ https://docs.x.ai/grok-bot/security-faq suggests exactly that. 3. After a test, check what’s still signed in. Look on the Bot’s computer and in your GitHub settings. Your Bots share one computer. Your code lives on others. Where your code lives Each Cloud Agent runs on its own machine, separate from your Bots’ computer. You can run many at once https://cursor.com/docs/cloud-agent , even with your laptop shut. In exchange, the agent can’t see your laptop. Code built from scratch lands on Origin, Cursor’s own code host, in early beta https://cursor.com/docs/origin . That’s fine while you test. When you want it on GitHub, ask a Bot to publish it. I asked QA Bot at the end, and it worked. I also wanted the code on my Mac, and told QA Bot: “Sorry I am not a tech person” . It sent me the whole app as a zip file. One question left: how big should your team be? How many Bots, and for how long Start with the smallest team that covers the job. So why not one big Bot for everything? My run gave me two reasons. A builder can’t be trusted to check its own work. QA caught two fixes the coder only claimed. And one Bot holding every rule forgets the oldest ones. Only the newest part of its memory reaches each reply. More Bots also means more replies, though. Anthropic measured its own multi-agent system at about 15 times the tokens https://www.anthropic.com/research/multi-agent-research-system of a chat. A busy scheduled task can use a week of usage in a day https://cursor.com/help/grok-bot/routines . Small tasks or big projects? Both worked for me. I used one Bot per small chore, and a team of three for a prototype. What about your next idea? I wondered whether my three Bots would mix the old project into the next. Each Bot keeps one long thread. Cursor’s staff confirmed there’s no fresh-session button yet https://forum.cursor.com/t/168333 . The cleaner option is Duplicate . According to SpaceXAI’s docs https://docs.x.ai/grok-bot/bots , the copy keeps the role, skills and routines. “It does not copy conversation history, learned memory, or chat attachments” . For a new idea, duplicate the team and give each copy the new scope. So how does all this compare with the tools you already use? Grok Bot next to Claude and ChatGPT - Claude Cowork gives each session its own sandbox https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview , deleted when the session ends. Claude picks its own helpers inside the task. As DataCamp https://www.datacamp.com/blog/grok-bot-vs-claude-cowork puts it, “In Grok Bot, you choose the workers. In Claude Cowork, Claude does” . - Claude Code runs on your machine when you work with it. Its scheduled tasks https://code.claude.com/docs/en/routines run in Anthropic’s cloud with your laptop closed. Its agent teams https://code.claude.com/docs/en/agent-teams , still experimental, let a lead session coordinate teammates. And every instruction lives in files you can read and edit. - ChatGPT Work and Codex give you a model picker https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex with a reasoning slider. The big labs now all ship an always-on personal agent. There’s OpenAI’s Dots https://openai.com/index/introducing-dots/ , Meta’s Muse https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ and Google’s Gemini Spark. Anthropic’s Claude Duet https://claude.com/resources/articles/claude-duet opened a US waitlist on Oct 21. An outside firm tested its security first. Each gives you one agent today, per Every’s comparison https://every.to/personal-agents-comparison . OpenAI writes that “we envision teams of dots working together” . Grok Bot is the one shipping a team now. That’s what I wish I’d known before my first Bot. Your turn I stopped my Grok Bot experiment when Opus 5.5 came out. I went back to my own setup in Claude Code. Here’s how I put it in my notes. I think I still need to control what is in the context and what are the instructions in skills, which is not the case with Grok Bot. So here’s where I draw the line. Grok Bot gets the work I can hand over without reading every instruction. Think of a prototype, a recurring chore or a first draft. Claude Code gets the work where I need to see and edit every instruction. This was a first test, and others run Grok Bot differently. I’ll keep testing. Pick one idea you’ve been carrying around, and create three Bots for it. Answer the PM’s questions, and read every word of the brief. Let QA say no as often as it needs to. Then ask someone, or something, outside the team to check the result. Your first build could start twenty minutes after you send the idea. To see what three Bots and four coding agents produce, the code is at The-AI-Thinker/quiet-desk https://github.com/The-AI-Thinker/quiet-desk . Go further Every source I used, grouped by who wrote it. The first group comes from SpaceXAI and Cursor, so read it as the vendor’s view. SpaceXAI and Cursor vendor - Grok Bot For Product Best Practices https://www.youtube.com/watch?v=x5OawRpQxII , Grok Bot’s product team on Cursor’s channel: a PM Bot writes the spec and loops in the designer itself video, 54 min . - Meet Grok Bot: Your Team of AI Agents https://www.youtube.com/watch?v=rkigdXf-52I , Amrita and Nate Umlauf: a QA Bot with a written test routine and screen recordings video . - Grok Bot 201 https://www.youtube.com/watch?v=rOVomdnWhjQ , Joseph Yang and Scott Lerner: three models on one design brief, and why each build needs its own repo video . - Grok Bot for PMs https://x.ai/bot/guides/grok-bot-for-pms , Kevin Niparko: specialist Bots over one catch-all assistant. - How I run multiple teams of Grok Bots https://x.ai/bot/guides/how-i-run-multiple-teams-of-grok-bots , Eric Zakariasson: one channel per project, and rules for staffing it. - Grok Bot for Engineering https://x.ai/bot/guides/grok-bot-for-engineering , Lingxi Li: before-and-after proof on every pull request. - Create and manage Bots https://docs.x.ai/grok-bot/bots , the official docs: roster size, roles and Duplicate. - Which AI models power Grok Bot? https://cursor.com/help/grok-bot/models , Cursor Help: how the model is chosen for you. Independent builders - We Built Grok Bot. Here Are Our 14 Best Bots https://www.youtube.com/watch?v=xZ5TEaleUdg , Peter Yang, with Grok Bot’s lead designer and engineer as guests: an engineering-lead Bot that never codes video . - 7 Grok Bot agents I use every day https://www.youtube.com/watch?v=QBmgF1kJSK4 , Claire Vo: Bots hired for one job each, with names to match video . - Grok Bot is more useful than I thought https://www.youtube.com/watch?v=RM8ZwhoJdeg , Brian Casel: how to stop managing your Bots in chat all day video . - A deep dive into Grok Bot https://flaviocopes.com/grok-bot/ , Flavio Copes: the limits and the comparisons, kept up to date. - Grok Bot vs Claude Code: I Run Both Every Day https://christopheralarcon.com/blog/grok-bot-vs-claude-code , Christopher Alarcon: a hands-on comparison, written about Claude Code’s interactive sessions. For the second check - Your job is to deliver code you have proven to work https://simonwillison.net/2025/Dec/18/code-proven-to-work/ , Simon Willison: what “done” should mean for code you didn’t write. - The lethal trifecta https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ , Simon Willison: the security read to forward to your IT team. - Agents Rule of Two https://ai.meta.com/blog/practical-ai-agent-security/ , Meta AI: give an agent at most two of the three risky powers in one session. - Cloud Agent logs claude-4.5-sonnet when picker is Grok 4.6 https://forum.cursor.com/t/169638 , Cursor’s forum: why your coders’ usage shows models you never picked, answered by Cursor’s staff. - Always-On Agents: Claude Duet vs Dots vs Grok Bot https://aimultiple.com/always-on-agents , AIMultiple: memory, logins and isolation, compared.