Can ChatGPT 6 Astra Clone Grokbot as a Local Desktop App?

Can ChatGPT 6 Astra one-shot a clone of Grokbot? That was the whole experiment. Grokbot is xAI's agent product: persistent bots, a remote computer per account, routines, skills, group chats. I wanted the same thing running locally on my Mac, powered by my ChatGPT subscription, and I wanted to see how far a single vibe prompt would get me.

Short version: two prompts, under two hours of build time, a working local VM, routines, skills, group chats, and a thumbnail I would actually publish. Also a lot of little bugs. Here is what actually happened on screen.

The setup

• ChatGPT desktop app with the Codex toggle on
• Astra set to extra-high reasoning
• An empty directory called GPTbot
• One screenshot of my existing Grokbot setup, with the toolbars expanded on my "Elon" bot
• A link to the Grokbot docs at docs.x.ai/grokbot

That is it. No spec document, no architecture diagram, no code.

The first prompt

I spoke the prompt rather than writing it. The gist:

Clone Grokbot one for one, end to end, but call it GPTbot. Power it with our OpenAI subscription so we can select Astra, Soul, Terra, Luna, and any of the GPT models on the account. Keep some discipline about not copying things outright. We do not want to step on SpaceX AI's toes, but a demonstrative clone should be fine.

Then the one unique constraint. Grokbot's remote computer runs on SpaceX servers. We are building locally, so we need some kind of Firecracker-style VM host on the Mac. Whether that is a daemon or an application running in the background, you decide. You are the technical one, not me. I just have the vision.

I attached the screenshot and hit enter.

Turn one: 1 hour 16 minutes

Astra planned, researched the docs, and built. When it finished, the build report listed what it had verified:

• Real ChatGPT sign-in
• Real responses from Astra, Soul, Terra, and Luna
• Persistent bots and group conversations
• Files, approvals, and editable skills
• A real local Linux VM using Apple Virtualization, with live desktop control
• Scheduled routines that keep running after the app quits
• Recorded demonstrations that produce reviewable skill drafts

It also added two notes that told me it understood the trade-offs. The model picker follows your account's Codex catalog, so there is no hardcoded model list. And local routines require your Mac to stay awake, which makes sense once you remember the compute is no longer in someone else's data center.

Then the honest part: "This is a private alpha. Full Grokbot parity remains unfinished." Provider integrations, voice, separate bot desktops, advanced recovery, and cloud/team features still needed work. It flagged those itself.

First launch

GPTbot was sitting in Applications. I opened it, and the first thing I saw was the first thing I did not like.

It had pre-created several bots and put a horizontal row of pinned options at the top of the sidebar. In Grokbot that top row is reserved space. It only fills in when you pin a bot. The sidebar normally just shows a list of your bots, and things like activity, routines, skills, and connectors live down in the bottom left as workspace options.

My screenshot caused this. I sent it with every toolbar expanded so Astra could see everything, and it took the layout literally. Screenshots communicate layout, not intent. Lesson learned.

Creating Sky

Bot creation worked cleanly. I hit plus, named the workspace Clearmud, and dictated the instructions for a YouTube agent: come up with new video ideas based on our format, our target audience, and what works, with a blend of what is trending, even though chasing every new trend is not our focus. I set the model to Astra on medium reasoning and named the bot Sky.

First test was cheap on purpose. I asked Sky what its purpose was, what its responsibilities were, and what it could do. It answered fast and it answered correctly, including the stance about not chasing trends. The system prompt propagated end to end on the first try.

The VM

I clicked Open Computer. The view came up, then appeared to freeze. I wondered whether I needed to hit Take Control, went full screen, hit it, and the desktop woke up. So it was not frozen, it was waiting on me. Browsing to Google and clearmud.ai worked.

Then I closed the browser and got a blank black desktop with no dock, no file manager, no terminal, no way to reopen anything. Bit buggy. It's a bit buggy. But the VM was real and it was running locally.

Routines, shared files, and the pulse

I created a routine called Check YouTube Comments: once a day, check unanswered YouTube comments and list them in Sky's chat. It saved. I could not find a way to force-run it immediately, which I wanted for testing.

Shared Files showed a workspace directory shared across bots. My guess at the time was one VM with a separate profile per bot.

The homepage surprised me. It showed "Sky needs your input," a pulse of what agents were waiting on. Grokbot does not have that. I would want to redesign the look of it, but the idea was a genuine addition the clone made on its own.

Sky's first real task

While I was poking around, Sky had been working on an editorial audit of the channel. Its strongest-angle finding was practical judgment: you build with AI, show what happens, and explain what deserves a place in your workflow. Fair.

Then the interesting line. It said it had reviewed our website tutorial index and six video companion articles because YouTube did not load reliably inside the VM, so it could not verify current views, thumbnails, or audience. It fell back to our blog as the source of truth and told me it had done so. That is exactly the behavior you want from an agent. It also told me where the configuration problem was.

Turn two: another 42 minutes

The second prompt was a long one. I asked for: GPTbot as one word with GPT uppercase, a simpler sidebar that starts with a plain list of teammates, optional sections, right-click menus on teammates, animated avatars during onboarding inspired by but not copying Grokbot, avatar generation using the subscription instead of picking a color, immediate control of the VM without the Take Control step, a dock inside the VM for browser, files, and terminal, one shared VM with a separate profile per teammate, a per-teammate model and reasoning slider like the ChatGPT app, editable souls and memory, and a look at the logs for why Sky's web scraping failed. Less is more. I also included Sky's response so it could see the failure.

Before sending it I checked usage. The first build had used 3 percent of my weekly quota on extra-high. That was the whole cost of turn one.

Forty-two minutes later, version two was installed with sign-in, conversations, and routines preserved. The build report explained the black screen: a ten-minute blanking timer in the VM, now disabled. Browser launching was repaired, and Sky's browser had successfully loaded the Clearmud YouTube channel.

Testing version two

The sidebar was clean. The reasoning slider in the top bar had a bug where some invisible element made it drag the whole window instead of the slider, though the same slider worked fine inside the edit-teammate panel. The avatar generator finished its job and produced a cloud with the face of Stannis Baratheon. Every bot had a unibrow. I laughed. Pinning a bot to the top removed it from its section, which was weird. Super weird.

Then I stopped looking at UI and gave the starter bots real work with zero instructions:

• Atlas: install the Last 30 Days skill by Matt Van Horn, then install Herder on the VM. It asked for approval to install. I approved once and it went.
• Craft: analyze the Clearmud YouTube channel and design a thumbnail using the rule of three. Me as the host, a ChatGPT logo, and the text "Meet GPTbot" with GPT highlighted.
• Scout: research Paddy Galloway, download transcripts of his last ten long-form videos, check the big shows he has guested on, and synthesize his approach to succeeding on YouTube in 2026 and beyond.

Sessions ran in parallel. Tasks got interrupted a couple of times with a "task was stopped, please check" message and no explanation, which is something I would want the app to self-resolve. But Craft delivered a thumbnail. I would use it. I would change the color to ChatGPT green and that is the only edit. It is going into A/B testing as one of three options. I gave it zero direction beyond the prompt and I was kind of blown away.

I also opened a group chat with Sky, Craft, and Scout to talk YouTube strategy. It worked, but the bots talked over each other with no clear distinction of who was replying to whom. For a third turn I would add @-mentions so I can tag one agent, the whole team, or the chat.

Last test: I opened the shared computer and watched my own live stream from inside the VM. Infinity mirror and beyond.

What I would take from this

• Vision-level prompting scaffolds an entire app. It does not polish one. The gaps were in UX details and reliability, never in breadth.
• The hard part of cloning a cloud product is infrastructure. Astra chose Apple Virtualization and shipped a working local Linux VM in one pass.
• Be careful with screenshots. Send only the layout you want copied, or say what matters in the image.
• Ask a new bot what its purpose is before giving it work. It is the cheapest check that instructions propagated.
• Read the fallbacks. When an agent says it used a different source than you asked for, the reason is where your configuration problems live.
• Budget over an hour for extra-high reasoning on a build this size, and treat the output as a private alpha.

Where this goes next

I could do a third turn, but the point was already made. Two prompts, less than two hours of build time, and a local Grokbot clone that signs in, runs a VM, schedules routines, installs skills, and designs thumbnails. It is still buggy. Nevertheless, it has never been easier to build, never been easier to clone, and never been easier to build something you can visualize.

Whether I keep maintaining GPTbot and publish the repo is up to you in the comments. OpenAI might eventually ask me to rename it, and the purpose here is not to build something for them. It is to inspire them to ship features like this inside ChatGPT or the Codex app. If you want the raw, uncut version of this build, it happened on the previous day's live stream.

I'm not an AI expert. I'm building in public and sharing what actually works. Drop a comment and let me know who you are, what you do, who your target audience is, and what your question is.

Watch the full video: https://www.youtube.com/watch?v=IXmruZvKsiU