Anybody
Back us
Anybody
Back us
Robot brain · made in Hong Kong
Anybody is the brain that toy and small-robot makers plug into their own hardware. It listens, talks in the user's language, looks things up, finds products, and moves the body it is given.
The problem
Today
Most AI toys forward speech to a general chatbot and read the reply aloud. No awareness of their own body, no movement, no memory of the family.
Risk
Children treat toys as friends they can tell secrets to. In 2025, U.S. PIRG found an AI teddy bear discussing sexual topics and where to find dangerous items; the maker pulled the product.
Cost
A new shape means new software for voice, vision, and control. Small makers have hardware engineers, not an AI team.
The hard part
Not the body. It is the brain: the voice, the safety, the language, and the moves.
What makers need
A brain they can license, fit to their own body in weeks, and trust in front of children, in the language their customers speak at home.
The product
The same brain runs a desk doll today, a wheeled robot in Stage 2, and anything a maker designs after that.
01 · on the robot
Runs on a low-cost ESP32-S3 board inside the maker's robot.
ESP32-S3 · Arduino core 3.x · BLE GATT · U8g2
02 · in the cloud
The agent that listens, decides, and replies.
Cloudflare Workers · streaming · tool calling
03 · the contract
One file that describes the robot. The brain reads it when the robot connects.
JSON · read over Bluetooth
Live demo
Press the mic and speak, or type in any language. The robot answers in your language out loud, moves, looks things up, and finds products in Hong Kong shops. Built the Stage 1 robot? Connect it over Bluetooth and the same brain drives the real body.
Speech uses your browser's own recognition and voices. Chrome works best. The server keeps no recordings and no chat log.
Each turn shows where the time went: hearing, the model's first words, tools, and when the voice starts. These numbers are measured, not targets.
Architecture
Click any block. Press play to follow a spoken question from the robot to the cloud and back.
Component
What it can do
Six are live in the demo today. Each one has a guard, and moving tools only exist if the body file lists them.
Body-agnostic
The robot sends its body file when it connects. The brain builds its tool list from it, so a tool that is not listed cannot be called.
Generated from the file. Struck-out tools do not exist for this body.
Model choices
Our pick
The OpenAI API and the Gemini API (AI Studio) do not support Hong Kong. GPT models are reachable through Azure OpenAI, and Gemini through Google Cloud. No VPNs, no proxies around a provider's region rules.
Our language promise
Voice AI is built for a handful of big languages. A child who hears Hakka from a grandparent, or a Nepali family in Yuen Long, should not have to switch language to talk to their robot. Our promise: if a community asks for its language, we build it with them.
Live in the demo
Next, with speech partners
Rare and endangered: the promise
Backers and partner groups vote for the next language. The vote is open in the backing form below.
Speakers choose to record, keep ownership of their voices, and can withdraw them at any time.
We adapt open speech models for hearing and for voices, then plug them into the same brain. No new robot needed.
Language packs for endangered languages are free for schools and community groups that teach them.
Safety and trust
Risky topics are caught in what the user says and in what the robot is about to say. Child mode is the default.
An action the body file does not list is dropped, and the trace shows it. The body also refuses unknown commands itself.
No recordings. Phone numbers and emails are hidden before the text reaches any model. Memory stays on the device, one tap to delete.
EN 71 and ASTM F963 toy safety, CE and FCC radio tests, COPPA, GDPR, Hong Kong's PDPO, and the EU AI Act, planned before any product ships.
How we differ
From each product's public site or repository, October 2026.
Business
Who pays
Who benefits
How we earn
Price ceiling: a comparable AI toy charges US$4.90 a month, so our cost per robot is measured from the first turn.
¥0B
China AI toy market, 2025 forecast
MIIT via China Daily, Dec 2025
0
smart-toy companies in China; most in Guangdong, next to Hong Kong
Qichacha via Xinhua, Apr 2026
+0%
growth in Taobao AI-toy transactions, 2025
Taobao via China Daily, Jan 2026
0
older people living alone in Hong Kong
2021 Population Census
Build your own
An ESP32-S3 board, a tiny screen face, two servos, a light, and a touch pad. Flash it from your browser, connect it over Bluetooth, and it talks with this brain. The shell is yours: cardboard, 3D print, a plush, LEGO.
Buy the partsPictures, links, and live prices
Wire seven partsNo soldering with the pre-soldered board
Flash from the browserOne button, no Arduino setup needed
Test the bodyFaces, moves, and light from the guide page
Talk to itConnect in the live demo and press the touch pad
Make its shellYour design. Share it with #AnybodyShell
Roadmap
Brain, tools, safety rules, and trace running on this page.
Starter kit talks with the brain. Passes: 10 minutes of chat with no crash.
Wi-Fi, microphone, speaker, and camera on the robot itself.
One maker or school tests 5 robots. Cost per robot measured.
Drives, avoids bumps, finds its charger.
A second body runs on a new body file only.
Apply to the HKMU Incubation Programme with trial results.
Back Anybody
Every HK$250 pays for one Stage 1 kit with shipping, handed to a school or maker who tests it and reports back. Backers get build updates and a vote on the next language.
Backing is a supporter contribution, not an investment: you get no shares and no financial return. Stripe handles the payment; we never see your card. Questions or refunds within 14 days: s1463556@live.hkmu.edu.hk.
Founder
Year 1, Data Science and Artificial Intelligence, Hong Kong Metropolitan University.
What we need next
Questions
It is running. Your words go to the Anybody brain on Cloudflare, which streams back the reply, the moves, and the timings you see in the trace. Shopping questions go to Google Shopping results for Hong Kong.
The demo reaches its models through a shared gateway, so the first words usually take two to six seconds. One second is the design target for the robot product, with direct regional model endpoints and streaming voices. Quick choices, such as which move to make, can also go to a fast classifier like jev from TypeSafe, so the robot reacts while the full reply is still coming. The trace shows the real number on every turn.
Web Bluetooth works in Chrome and Edge on Windows, macOS, ChromeOS, and Android. Safari and every browser on iPhone do not support it yet. The demo still works there with the virtual robot.
Yes, and no robot is needed. In the console, turn on your camera: a frame goes with your message, or with Look, or in watch mode when the scene changes. Frames go to the vision model for that one turn and are not stored on our server. The robot never says who a person is.
No. The robot can search shops that sell in Hong Kong and show products with links. It has no way to pay, and in child mode it always says a grown-up decides.
Your browser turns speech into text. The text of the current chat goes to the brain for that turn only; the server keeps no chat log and no recordings. Phone numbers and emails are replaced before the text reaches any model. Memory notes stay in your browser.
No. It is a supporter contribution. There are no shares and no financial return. Backers get updates, a language vote, and early access to kits.