Quick answer
An AI desk robot assistant is a small physical robot that listens for a wake word, answers questions, checks your calendar and email, and talks back, with a frontier AI model doing the thinking. In his latest video, Mark Kashef shows how he bought a $99 M5Stack desk robot, plugged it into his Mac and had Claude Fable 5.2 plan and build the first versions of new software and firmware, before switching to Opus 5.5 to finish the job over 20 versions. His point is simple: these models can work on real hardware, not just another app.
Key takeaways
- Hardware is now fair game. If a device plugs into your computer and has public docs, a frontier model can help you reprogramme it.
- The prompt did the heavy lifting. Mark’s eight-part brief covered research, backup, goals, voice, wake word, response style, restrictions and testing.
- Safety was designed in. The robot can read and draft email but is not allowed to send it, and recordings and camera checks stay on his Mac.
- It still broke a lot. Lag, missed wake words and misfiring tools took many rounds of fixes.
- Use the expensive model to plan. Mark used Fable 5.2 for planning and the early builds, then Opus 5.5 to iterate.
In this guide
- What is an AI desk robot assistant?
- What can Mark’s robot actually do?
- How did Claude Fable 5.2 rebuild a $99 robot?
- What was in the eight-part build prompt?
- How does the cloned voice work?
- What went wrong across 20 versions?
- How does one voice request flow through the system?
- What can a small business learn from this?
- Should you build one yourself?
- Watch: the source video
- Frequently asked questions
- Related posts
Most new-model demos show a game, a website or an app you will never use again. Mark Kashef took a different route with Claude Fable 5.2 and pointed it at a real object on his desk. This article breaks down what he built, how he prompted it and what went wrong, then pulls out the lessons that matter even if you never build a robot. If you want our wider verdict on the model itself, read our Claude Fable 5.2 review.
What is an AI desk robot assistant?
An AI desk robot assistant is a voice assistant with a body. Instead of talking to your phone, you talk to a small robot that sits on your desk. It shows faces on a screen, moves, reacts when you touch or pick it up, and replies out loud.
Mark’s robot is called Lil Token. It weighs about as much as a phone. The clever part is not the hardware, which was bought off the shelf, but the software Claude wrote to connect it to his AI subscription and his everyday tools.
What can Mark’s robot actually do?
Based on Mark’s demo, Lil Token can:
- Answer questions and come up with its own replies, using his Claude subscription rather than scripted lines
- Research questions on the web (it shows a small flame icon when it is using Firecrawl, the research tool he uses)
- Check his calendar and read his real inbox, then draft replies without sending them (a Google icon shows when it is using Google tools)
- Connect to hundreds of other apps through Composio, plus Slack, Stripe and various command-line tools
- Reply in a clone of Mark’s own voice
- React to being touched, picked up and put down
- Take a quick photo every 20 seconds to notice when he is back at his desk, then welcome him and tell him what is next in his day
- Hear music, guess the genre and dance to it
- Tell him when a long-running Claude task on his Mac has finished
Mark says the recordings and camera checks stay on his Mac, and that a few small local models handle simple admin tasks. He also mentions a cloud “second brain” that keeps common requests on file so it can answer faster.
⚠️ Warning: A robot that takes a photo every 20 seconds is a camera in your room. If you build anything similar, keep images local, delete them quickly, and tell anyone who shares the space. In an office, check your workplace and data protection rules first.
How did Claude Fable 5.2 rebuild a $99 robot?
Mark started by asking Claude to research the three best personal robots that it could fully reprogramme from a computer. He chose the M5Stack robot (the description calls it a StackChan) as the best balance of price and features. Out of the box it came with a phone app, remote control and smart home features, but he did not want his conversations and camera snapshots going to someone else’s cloud.
So he had Fable take stock of the hardware, which included:
| Component | What it is used for in Mark’s build |
|---|---|
| Touch screen | Faces and status icons for each emotion or tool |
| Camera | Presence check every 20 seconds, kept on his Mac |
| Two microphones | Wake word, voice requests and hearing music |
| Speaker | Replies in his cloned voice |
| Motion sensors | Knowing when it is picked up or put down |
| Lights and motors | Reactions and dancing |
| Wi-Fi chip | Talking to his Mac and online tools |
Then came the important safety step: before wiping anything, Fable made a full backup of the original software and firmware. After that it rebuilt the software and firmware from scratch. Mark gave it the product page and manual so it could tell him what could and could not be changed.
💡 Tip: If you ever let an AI tool change a device, insist on a full backup and a written “how to restore” note first. It turns a costly mistake into a five-minute fix.
What was in the eight-part build prompt?
Mark’s build prompt had eight parts, and you can reuse the structure for almost any hardware or automation project. Here is our plain-English summary:
- Research and backup. Search the exact make and model, read the maker’s docs, product page and open-source firmware projects, check what owners say in forums, break down every component, and back everything up. Mark also let Claude use iPhone mirroring to click through the phone app and connect the robot to Wi-Fi.
- The goal. Describe the dream outcome: swap the robot’s built-in model (Qwen) for Opus 5.5 through his Claude subscription using the Claude SDK, reply in his cloned voice through Fish Audio, and connect research, Gmail, calendar and other apps.
- Faces. Define the screens and expressions for each reaction.
- Wake word. “Hey Lil Token” starts listening, and it keeps listening for about 10 to 15 seconds so you can finish your sentence.
- Response style. How short and to the point replies should be.
- Restrictions. What the robot must never do, such as sending emails on his behalf, in case of mistakes or made-up answers.
- Troubleshooting. Ways to log and trace every error, including frequent screenshots while testing features like music detection.
- Simulated testing. Use the Mac’s speakers to hold synthetic conversations with the robot while Claude listens and records what breaks.
Our related guide on using AI to plan a simple electronics gadget covers the planning side if you are starting from parts rather than a ready-made robot.
🔎 Unconfirmed / rumour: Mark says using a Claude subscription through the Claude SDK for personal projects should be fine under the terms of service, but he also says to do your own checks. We have not verified this. Read Anthropic’s current terms before you connect a subscription to any device, and do not use a personal plan for a commercial product.
How does the cloned voice work?
Mark cloned his voice by giving Fish Audio about 15 of his earlier YouTube videos. When the robot replies, the text goes to Fish Audio’s API with the ID of his cloned voice, and the audio comes back to the robot’s speaker.
He used Fish Audio’s newer expressive model (he calls it Drama 3), which accepts bracketed directions such as startled, happy or excited. Claude read the API docs and added a direction to each pre-written line, so a phrase like “Hey, easy, I’m getting dizzy” sounds alarmed rather than flat. If you want to write directions like this yourself, our guide on directing AI voice acting with plain-English script notes walks through it.
⚠️ Warning: Only clone a voice you have the right to use, which in practice means your own or one you have written permission for. Keep the voice ID private, as anyone with it could make “you” say anything.
Fish Audio sponsors this part of Mark’s video, so treat his enthusiasm for it as a sponsor’s view and compare it with other voice tools before you choose.
What went wrong across 20 versions?
Even with a detailed plan, the first versions were rough. Mark shows a “wall of shame” of bugs. The main ones were:
| Problem | What happened | How it was tackled |
|---|---|---|
| Lag | A long gap between speaking and the reply | Repeated testing and tracing each step |
| Touch and speech | The robot ignored touches and some requests | Fixes across many versions |
| Tool confusion | It could not tell whether he was talking about a tool or asking it to use one | Test sessions comparing tools requested with tools that actually fired |
| Mumbled wake word | Missed “Lil Token” when he mumbled | Training it on many ways of saying the name |
| Endless dancing | Its own motor noise drowned out everything, so it never stopped | Adding stop phrases, which it also tended to mishear |
| Odd visuals | Green bands across the face at one point | Fixed in later versions |
It took 20 versions in total. Mark used Fable 5.2 to plan and build the first few, then switched to Opus 5.5 to keep iterating, partly because Fable’s usage limits are tight. Because every session left a trace, the model could read the logs and improve the next version each time.
Want the practical side? Our step-by-step guides show you how to let an AI assistant read your email without sending anything and how to test a voice assistant with simulated conversations.
How does one voice request flow through the system?
A single request, such as “What’s on my calendar?”, takes a short round trip.
- The robot hears the wake word and records your question.
- Claude turns the speech into a request and decides which tool to use (calendar, research, email or another app).
- The tool returns the answer and Claude writes a short reply.
- Fish Audio turns the reply into speech in Mark’s cloned voice.
- The robot speaks, showing an icon for the tool it used. While it is thinking, a capital “T” appears in the top-right corner.
Showing what the assistant is doing, with a simple icon, is a small touch that builds trust. We cover the same idea for software in How to Show Users What Your AI Agent Is Doing, Without Technical Jargon.
The faces were generated with the Codex CLI and matched to emotions and physical states, such as “picked up” or “put down”. For music, Fable trained it to recognise patterns in genres such as jazz, classical and Middle Eastern music so it could dance in a matching style.
What can a small business learn from this?
Mark is clear that this was an experiment rather than a business project, although he likes hearing his day without picking up his phone. The useful lessons travel well:
- Write the brief before you build. His eight-part prompt is a solid template for any automation: research, backup, goal, behaviour, limits, testing.
- Draft, don’t send. Giving an assistant read access but no power to act on your behalf is a sensible default. Our guide on letting an AI assistant read your email without sending anything shows how.
- Test without being in the room. Simulated conversations found problems while he got on with other work. See how to test a voice assistant with simulated conversations.
- Keep personal data local where you can. Wiping the maker’s cloud features was a privacy choice, not a technical one.
- Match the model to the job. Plan with the strongest model, iterate with a cheaper one. Our Claude model guide explains the trade-offs.
If you are thinking about AI as part of how you run your business day to day, Peter Yang’s personal AI system is a good software-only comparison, and our Claude Fable 5.2 business ideas article covers ways to make money with the model.
Should you build one yourself?
Only if you enjoy tinkering and are happy for it to take time. Mark is not a hardware person, and he still needed 20 versions and plenty of patience. Ask yourself:
- Do you have a clear daily job for it, such as reading out your calendar each morning?
- Are you comfortable giving an AI read access to your inbox and calendar?
- Can you keep camera and microphone data on your own machine?
- Will you check the terms of the AI subscription and voice service you plan to use?
Mark has open-sourced the project, setup guide, architecture, faces and prompts as a free kit linked from his video description. He says he has a 3D printer on the way and plans to use Fable to work out what to do with it next.
Watch: the source video
What this video covers: Mark Kashef turns a $99 M5Stack desk robot into a voice assistant called Lil Token using Claude Fable 5.2 and Opus 5.5, explaining his eight-part build prompt, voice cloning, the bugs he fixed and how a request flows through the system.
Video: Mark Kashef (@Mark_Kashef), “Claude Fable 5.2 Can Do So Much More Than Build Apps”, published 13 November 2026. The voice segment is sponsored by Fish Audio, and Mark’s Fish Audio link is an affiliate link; he says he is not affiliated with M5Stack, the robot maker. This summary and the opinions in it are our own; build details and results are Mark’s own account.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Let an AI Assistant Read Your Email Without Sending Anything — set up read-and-draft-only access so your assistant can never send on your behalf.
- How to Test a Voice Assistant With Simulated Conversations — let your computer talk to your voice assistant and log every failure while you get on with your day.
Next step: pick one small daily job you would hand to a voice assistant, write a one-page brief using Mark’s eight parts, and decide what it must never do before you build anything.
Frequently asked questions
What robot did Mark Kashef use?
A $99 desk robot from M5Stack, called a StackChan in his video description. He chose it after asking Claude to research robots that could be fully reprogrammed from a computer, and says he is not affiliated with the company.
Did Claude Fable 5.2 build the whole thing?
No. Mark used Fable 5.2 to plan and build the first versions, then Opus 5.5 to iterate to the final version, 20 versions in all.
Can the robot send emails?
No. Mark deliberately set it up to read his inbox and draft replies, but not to send them, in case of bugs or made-up answers.
How did he clone his voice?
He fed about 15 of his earlier YouTube videos into Fish Audio, which sponsors the voice segment of the video, and the robot uses that voice ID for its replies.
Do I need coding skills to try this?
Mark says he is not a hardware person and had never programmed a robot, but he does have a lot of experience directing AI tools. Expect a learning curve and plenty of testing.
Is the project available to download?
Yes. Mark links a free kit with the open-source project, setup guide, architecture, faces and prompts in his video description.
Related posts
- Claude Fable 5.2 Review: Is It Really the World’s Best AI Model?
- Which Claude Model Should You Use? Haiku, Sonnet, Opus or Fable
- How to Use AI to Plan a Simple Electronics Gadget: Parts List, Code and 3D-Printed Case

[…] guide follows on from our article AI Desk Robot Assistant: How Mark Kashef Rebuilt a $99 Robot With Claude Fable 5.2. Mark Kashef’s robot reads his real inbox and drafts replies, but he deliberately stopped it from […]