Most AI coding tools throw everything at you. Tool calls, file trees, shimmering “thinking” text, model pickers, terminal output. It’s a lot. Then along comes an app that hides all of it and just… talks to you. I stumbled onto this one through a walkthrough from AI creator Matthew Berman, and I couldn’t stop watching. He digs into a new app called Grockbot, and honestly, I think it points at where this whole space is heading.
Here’s the quick version. Grockbot is an agentic system built for everyone, not just developers. The original poster describes it feeling like you’re just texting your agent. Threads on the left, chat on the right. That’s it. No code, no directory structure, no jargon. And under that simple skin, it’s still seriously powerful.
🧭 The old way vs the new way
This is the contrast that got me. In tools like Cursor, Codex, or Claude Code, every thread is basically the same agent, just split up by topic. One thread for a bug fix, another for a PR. You also see everything: the files being edited, the searches, the tool calls, the whole path to the answer.
Grockbot flips both ideas.
- Every thread is its own individual agent, not the same agent sorted by topic. The creator says it really makes each one feel like a distinct coworker.
- Nothing technical is shown. No tool calls, no code, no file reads. Just point A to point B.
- There’s no model picker. You can’t pick between Grok, GPT, or anything else. You don’t even know which model is running. The person who shared it thinks most people genuinely don’t care, and I agree.
That’s the bet here: strip out what confuses non-technical folks and keep the power humming quietly underneath.
💻 The cloud computer trick
This is the part where I sat up. When the author asked his shopping agent to “show me live on Amazon,” a little computer icon appeared in the corner. Click it, and you see the agent spun up its own full operating system. A real Linux desktop, file manager and all. Every single bot you create gets its own fresh computer.
And here’s the clever bit he points out: even though each agent gets a separate environment, they share authentication. Log into Amazon once, and every future agent already has you logged in. You can also grab the mouse and do work yourself inside that cloud machine anytime.
It’s a hybrid too. The creator asked it “how many folders are on my desktop” and it counted his local files just as easily as it browsed the cloud. Local plus cloud, sliding between them without friction.
🤝 Agents that talk to each other
This feature is genuinely new to me. The original poster set up a “chief of staff” agent as his main point of contact, an idea he credits to a suggestion from the Cursor team. When he asked it something bigger, like how many emails were safe to archive, the chief of staff messaged his dedicated email agent to handle it.
You can click into that conversation and read the two agents talking. Even better, those agent-to-agent chats persist forever, quietly building up context you can reuse later. He’d never seen that in another app, and neither have I.
🛠️ Teaching it your own tasks
The “teach a task” button is my favorite bit. The author wanted a weekly camera price check, so he opened the agent’s own computer, grabbed a product URL, pasted it into a spreadsheet, copied the price, and hit stop. Grockbot watched the whole demonstration and turned it into a reusable skill on its own. No coding. It literally learned by watching him work once.
You can also add prebuilt skills. He installed a “humanizer” skill to make AI writing sound less robotic, and when it wasn’t in the marketplace, Grockbot just pulled it straight from the repo as a private skill.
Want to try it yourself? Here’s the flow the creator walks through:
- Hit the plus button and either pick an existing agent or create a new bot.
- Start typing your goal. It asks smart follow-up questions to figure out what you actually want.
- Add plugins (Gmail, Drive, Calendar, Slack, Notion, Box and more) with one click to expand what it can reach.
- Set up routines by right-clicking an agent, editing its profile, then creating a scheduled task in plain language. His “computer cleanup” bot runs every Monday at 8am.
- Use “teach a task” to record yourself doing something once and let it build the skill.
🤔 The honest catches
The person who posted it doesn’t oversell it, which I appreciate. A couple of real snags:
- Routines are buried behind a weird menu path (edit profile, then a back button). Not obvious.
- Having a separate app just for knowledge work feels awkward when your code and context already live in Cursor. The line between “do this in Grockbot” or “do this in Cursor” gets blurry. He points at how OpenAI merged Codex and ChatGPT into one app and wonders if splitting them is the right call.
His dream, and mine, is one place where you drop in a task and it just knows how much detail to show you.
If you want to see the cloud OS spinning up and the agents messaging each other in real time, watch Matthew Berman’s full walkthrough. It’s worth it just for the “teach a task” moment.