How to Use Claude Fable 5.1 Inside Grok Bot: 3 Methods

You do not have to pick between Grok Bot and Claude. You can run Claude's Fable 5.1 inside your Grok Bot organization, and if you route the work through a CLI on the remote machine, it barely touches your usage.

That last part is the actual reason to do this. In the video, a 20-minute autonomous website build finished and the usage meter read the same number it did before the build started.

Here are the three routes, ranked the way Marcelo ranks them on screen.

Route 1: Shepherd bot template via Herder (the favorite)

This is the one demoed at length, and it gets declared the favorite in the first fifteen seconds, before any of the three are shown.

The shape of it: a Shepherd bot installs Herder on the Grok Bot remote PC. Herder is what holds always-on agents running in the CLI, so you can spin up as many sessions as you like, and it has control over every one of your CLIs. In the demo, the engineering channel has Grok, Claude, and Codex each configured as its own separate session on that remote machine.

Open the Claude session and it looks exactly like a local one. As the video puts it, it is just as if you spun up Claude Code in the CLI locally, but it is all happening on the remote Grok Bot computer.

One note worth respecting: the video deliberately does not teach Herder. "If you're unaware of what Herder is, I highly recommend you check it out. I'm not here to explain Herder." Treat Herder as a prerequisite you go learn separately, not as part of this tutorial.

Attribution matters here. The Shepherd prompt that built all of this came from a community post by Lauren, aka Potato, on X. Marcelo asked Shepherd to look at X and check out Lauren, misspelled the name, and Shepherd still found the Bot village post with the copy-paste prompt inside it. That prompt is the source. Credit stays with Lauren.

Route 2: Peekaboo on your Mac invoking Claude in the CLI

Set up Peekaboo on your Mac to invoke Claude in the CLI, where it fires a claude -p command.

This route is described on screen with its own prompt, but it is not demonstrated in the video. Take it as a stated option, not a tested one.

Route 3: Claude Code installed on the remote computer

Install Claude CLI, Claude Code, directly on your remote computer. Authenticate your subscription. Then dedicate one bot to it, name it after the responsibility, and route every Claude-bound prompt through that bot. In the video that bot is simply named Claude.

"It's as simple as that." Also described, also not demonstrated.

You are not supposed to install any of this by hand

This is the part that makes all three routes usable in an afternoon. Each route gets its own prompt on screen, and the instruction is identical for all three: copy it or screenshot your screen, feed it to your bot, and it will guide you along the process.

The Shepherd bot template is linked, and it is also available in the YouTube description as a fallback.

That is exactly how the demo setup happened. One conversational instruction, and the bot installed Herder on the remote PC and staffed a whole roster on its own: leads bot, motion bot, engineer, drafters. "All hands off. I didn't have to do anything."

The usage proof

The demo is structured around a single number.

Before the build, the usage meter reads 4%. Honest caveat, stated in the video and worth keeping: that reading came right after a usage reset.

Then one prompt goes out. Create a new session, build a new website for HVAC companies in the greater Las Vegas area, use Claude's Fable 5.1. That is the entire creative direction.

It runs for 20 minutes, unsupervised, and comes back with a finished site. The on-camera reaction is unedited and slightly incoherent, which is the point: "Look, it built us Wow. Very impressive. I'm loving the style here. And we gave it very little creative direction like, I, I, ah, this is such a great little workaround."

Then the meter: "Look at our usage. It's the same exact usage here."

That is the whole argument. If you are getting rate limited inside Grok Bot, this is the workaround.

The bug that did not get edited out

Shepherd finished the build and never reported back. No completion notification.

"Looks like we do have to fix a little issue where Shepherd didn't come back and let us know that it was done. But nevertheless, it did the work. It started the work, exactly as intended."

Practical takeaway: until notifications are reliable, check the work yourself. A silent agent is not necessarily a stalled agent.

What to actually do

• If you are rate limited, take route one. Shepherd plus Herder on the remote PC is the explicit recommendation.
• Do not install Herder by hand. Screenshot the route prompt, hand it to your bot, follow its guidance.
• Go learn Herder on its own time. It is a prerequisite, not a step.
• Run one CLI session per tool. Grok, Claude, and Codex each get their own always-on session.
• If you take route three, dedicate a single bot to it and route every Claude prompt through that one bot.
• Check your usage meter before and after a long build. That is the test that tells you whether this is working for you.
• Give long-running agents minimal direction and let them run. One sentence bought 20 minutes of work worth keeping.
• Verify completion yourself until the notification issue is fixed.

Caveats worth stating plainly

The 4% reading follows a reset. The completion-notification bug is unresolved. Routes two and three are described on screen but never run in this video. Nothing else in the demo is quantified, so do not read cost savings or time-saved figures into it that were not measured.

Three routes. One favorite. One usage meter that did not move.

Want a video on your specific setup? Comment with who you are, what you do, who your target audience is, and your question.

"I'm not an AI expert. I'm building in public and sharing what actually works."

My name's Marcelo. This is Clear Mud and clarity matters.

Watch the full walkthrough: https://www.youtube.com/watch?v=ZZ8c5oA3YTM