Breaking down the Grok Bot: What a cloud Agent adds
OpenClaw is open source and self-hosted; the Grok Bot is the opposite: closed source, hosted in the cloud. You cannot see its source code, but the official documentation is very detailed. This site has a Grok Bot course that teaches you to use it; this section looks at the design choices it makes as an Agent and how it differs from the myagent you wrote.
The content below comes only from the Grok Bot's official documentation; for anything the documentation does not cover, we explicitly mark it as "not disclosed."
One-sentence summary
The description from the official release notes: the Bot is "a persistent AI teammate that works on a persistent cloud computer, with messaging, approvals, connected apps, and routines." Each Bot has a name and responsibilities, and works on a cloud computer with a browser, file system, and terminal.
Seven building blocks
Look it upBlocks
Click one to see.
Compared with myagent, what does the cloud add
| myagent (local) | Grok Bot (cloud) | |
|---|---|---|
| Where it runs | Your computer; it only exists when you are there | A persistent cloud computer, always on |
| Who vets actions | Asks you y/N every time | The review model judges first, and only sends you an approval card when unsure |
| Isolation | None (the exercises add a working-directory restriction) | A hardware-isolated virtual machine per user |
| Memory | Conversation saved to file + summary | The Bot proactively records facts and logs, which can be viewed and deleted |
| How it gets woken up | You type in the terminal | Schedules, messages, external events |
| Which model it uses | You choose | Not disclosed in official docs; may differ per request |
The two points most worth learning:
- A review model replaces part of manual confirmation. Popping a y/N every time is safe but annoying. Grok Bot and OpenClaw's auto mode both use the same idea: first let a model judge whether the action is safe, and only ask a human when unsure.
- Put the Agent in a computer that "doesn't hurt if it breaks." The sandbox discussed in the previous chapter, Grok Bot turns directly into one virtual machine per user.
Hands-on in the Grok Bot course
| Which part you want to dig into | The hands-on page in the Grok Bot course |
|---|---|
| Conversation, interruption, stopping | Conversation basics |
| Approvals | Approvals |
| Memory | Memory |
| Skills | Skills |
| Connected apps and MCP | Connect apps |
| Routines | Routines |
| Multiple Bots | Multi-Bot collaboration |
Quiz timeWhat problem does Grok Bot's Auto Review solve?
Check yourself
- I can name several things a cloud Agent has over a local Agent
- I can tell which parts of the Grok Bot documentation are public and which are not
- I know which page of the Grok Bot course to go hands-on with
Next section: Breaking down open-source coding Agents: Codex, OpenCode, goose.