yuanyijie

Learn Grok Bot — Development skill for Claude Code

Development community

Learn Grok Bot's desktop agent architecture in 16 lessons: Electron process isolation, trusted RPC, turn loop, streaming transcript, inference router (Cursor/Claude Code/Codex/OpenRouter), sandbox, MC.

How to install Learn Grok Bot

This entry records only its repository, not the path inside it, so there is no exact command to give. Open yuanyijie/learn-grok-bot and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Learn Grok Bot does

Learn Grok Bot's desktop agent architecture in 16 lessons: Electron process isolation, trusted RPC, turn loop, streaming transcript, inference router (Cursor/Claude Code/Codex/OpenRouter), sandbox, MCP, and automations. Mock-first harness course. Not official.

Alternatives in Development

  • Microsandbox — Open-source self-hosted MicroVM sandboxes with sub-200ms startup, hardware-level isolation via libkrun 5k ★
  • Magnitude — Open source inference server your agent sets up for you 1.5k ★
  • Android Rpc — Android Debug RPC 1.3k ★

README

[English](./README.md) | [中文](./README.zh.md)

Learn Grok Bot — An AI Employee With Its Own Computer

Grok Bot is not a chatbot. It is not a coding CLI with a window.

It is a **desktop AI employee**: a named persona, a Linux computer of its own, and one voice (`SendMessage`). Who wakes it changes the rules. Dangerous work goes through a gate, then a human.

![Product graph](images/product.svg)

Each lesson does two jobs: **name the product object**, then **show how the reconstructed tree enforces it**. Objects without code are a story. Processes without objects are a roster.

  • PRODUCT.md — five objects, the wake table, one full task
  • ARCHITECTURE.md — product graph + runtime graph + product verbs = which RPC table

The mindshift

A terminal coding agent owns your repo. A Grok Bot employee owns **another computer**.

The polished UI is the least trusted process in the building. It paints `SendMessage` as bubbles. It must never run a shell, read a secret store, or leak the model's inner monologue into the chat.

Product fact If you invert it
The user hired a roster employee You shipped a completion
Plain model text is invisible Drafts and tool plans flash in the chat
"My computer" is the shared box rm runs at home
A [routine] may stay silent Cron spam: "(no change.)"
A widget ends the turn The employee keeps talking while waiting

Renderer / preload / main / coordinator / host / box are load-bearing walls. The walls are not the product.


16 lessons, by object

**Each lesson = product invariant + implementation dissection + runnable teaching code.** The motto is a reminder. The body must say which process owns the object, and whether prompt, middleware, or a hard gate nails the invariant.

Stage 1 · Product objects

**s01**   *"You hire an employee, not a completion"* — roster, persona, `[first run]`

**s02**   *"Plain text is a monologue; SendMessage is