Grokdir

grokdir

How the Grok Bot interface works

xAI designed Grok Bot around persistent bots, not chat sessions: a roster, presence, each bot's own computer, and routines that start without a prompt.

xAI wrote up how it designed Grok Bot. Most AI products are built around a chat you start, watch, and close. Grok Bot is built around workers that keep going after you leave.

This is a short map of that design, useful if you are adding a share from grokdir and want to know what you are looking at.

Five things the product exposes

AI tools have piled up a long list of nouns — sessions, models, context windows, memories, system prompts, projects, skills, connectors, sandboxes. xAI’s view is that showing all of them asks people to understand more than they need.

They kept five:

  • Bots — persistent agents with a name, memory, runtime, and tools
  • Chats — how you talk to a bot
  • Prompts — context or instructions, used once, saved as skills, or fired as routines
  • Tools — how a bot reads information and acts, including software, APIs, the shell, and computer use
  • Artifacts — the files, designs, code, and other lasting work a bot makes or changes

The app is organized around bots, not around a chat history.

A roster, not a history

The sidebar is a list of bots. Each has a name, an avatar, and a title. It keeps the conversations you had with it. It has its own computer. Tomorrow you open the same bot, not a new thread.

The avatar is meant to answer three things at a glance: who this is, what it is doing, and how much you need to look at. State lives on that same avatar — idle, thinking, working, waiting, blocked, done. Hover if you want the current action.

The bot’s computer

Each bot has a computer of its own for the browser, files, and software. A title-bar icon turns purple while the computer is in use. Opening it pins a side panel so you can watch without leaving the chat. If the bot needs you, you can take the computer full screen, do the step, and hand it back.

How answers and work show up

Inline cards and widgets let a bot use prose when prose fits and structured UI when it does not. Creating a routine, changing a setting, or messaging another bot can show up in the same transcript.

When people have several bots, some make one of them a coordinator and give direction there instead of checking every bot. Tools and skills sit at the account, because many bots need the same browser, documents, or mail. Memory and routines stay on the bot. Group chats share project context without merging those memories.

Most agent products wait for a prompt. Routines are standing work on a clock or an event — an industry watch, a briefing every morning. A prompt can still open a session. So can a schedule, an event, or another bot.

What they took out

xAI capped accounts at about 50 bots and group chats at six, and cut extra chrome around the computer view. The test for each of those cuts was whether it helped someone delegate.

The full write-up is on xAI.