Vision
- Turn to look at you when you walk inFace tracking
- Recognize who’s at the deskFace recognition
- Describe whatever you hold upVision-language models
Physical AI, for software people
Give your code a face.
Deskman is a compact robot built for developers who want to learn physical AI. Talk to it, script it, swap its parts, rewrite its brain. Deskman is the physical AI platform made for developers who learn by experimenting.
Hellomy name isDeskman
If you can code, you're ready.
Deskman teaches you the rest.
Skills you need
- Use the command line
- Cursor or Claude Code (a dev environment)
- Comfortable with AI cloud accounts (OpenAI)
Skills you'll learn
- How to control motion: servos and parts
- How to do face tracking from a camera feed
- How to develop and interact with a voice AI agent
Teach Deskman tricks. Learn real skills.
Every project teaches a skill you can use anywhere. Here are a few to start with.
Voice and AI
- Answer questions with a local LLMOn-device AI
- Nod along and react while you talkConversational AI + motion
- Translate a conversation in real timeSpeech-to-speech
Expression and motion
- Show emotions on its faceAnimation + state machines
- Glance around when a message arrivesMotion control
- Follow the speaker in a meeting roomSound localization
Just for fun
- Recognize the ingredients on your counter and suggest a recipeObject detection + LLMs
- Nod back when you give it a thumbs-upGesture recognition
Get Deskman
Shipping time: 1–2 weeks · Ships from CA, USA
Open software.
Deskman isn't a sealed consumer assistant — it's designed to be modified. The software runs on Linux and Python, and the hardware is accessible, so you can:
- Change how it behaves
- Replace models with your own
- Modify the motion system
- Add sensors
- Start your own robot from Deskman as the base
- Remix the body — it's 3D printed, so it's easy to change
$ ssh deskman@deskman.local
deskman@deskman:~/deskman$ cd talk
deskman@deskman:~/deskman/talk$ ./talk.py
Loading whisper...
Loaded on GPU in 5.7 sec
Loading kokoro...
Loaded on GPU in 10.0 sec, voice bm_fable
Loading text model server...
Loaded in 7.5 sec on GPU. Model: gemma-4-e2b
Starting text model inference...
Started in 4.5 sec
Robot ready and listening. Say "Deskman" to talk.Three motors. One camera. Many possibilities.
Face tracking, person and object detection, and motion work
out of the box.
Then you can reprogram all of it.
Look left & right
Turns side to side to follow people around the room.
Look up & down
Tilts to nod, look up at you, or glance down at your desk.
Privacy cap
Lowers over the camera, so you always know when it’s watching.
Expressions
Draws a new face on its screen to react to what you're saying.
What's inside
NVIDIA Jetson Orin Nano, or Raspberry Pi 5 4GB
7 inch touchscreen
1080p front camera
Speaker + audio amplifier
Microphone + audio card
Three servos + control board
Battery (5,800mAh Li-ion) + charger
Battery level sensor
3D-printed body, head & hat
Use the models you want. Modify the code. Build something weird.
- What does it run on?
- Linux and Python, with CUDA accelerating speech-to-text, the LLM, and text-to-speech.
- Which models can I use?
- Modern open-source speech-to-text, text-to-speech, computer vision, and local language models. Swap components as new models arrive.
- Am I locked into a proprietary AI service?
- No. You choose the models, and you can change them whenever you want.
- Can I change the code?
- Yes. It's open source on GitHub, so you can read it, modify it, and make it your own.
Put a robot on your desk.
SSH in. Start learning.
Deskman arrives assembled and ready to develop on.
Not ready yet? Get Deskman news and updates.