OpenToys, an open-source platform for building AI-powered toy companions, was released in March 2026 by akdeb, the founder of Elato AI. The project enables developers, makers, and educators to construct Yoto-like interactive toys using an ESP32-S3 microcontroller paired with an Apple Silicon Mac — running the entire AI stack locally with no cloud dependency. The decision to open-source the commercial Elato AI product stack came directly from parental feedback: after akdeb initially released portions of the codebase, parents voiced strong objections to their children's conversations being transmitted to remote servers. OpenToys eliminates cloud transmission as a structural matter rather than a policy one.
Whisper Turbo ASR handles speech recognition. Text-to-speech runs through Qwen3-TTS — covering 10 or more languages including English, Chinese, Spanish, French, and Japanese — and Chatterbox-turbo. Any LLM available through the mlx-community on Hugging Face, including Qwen3, Llama, and Mistral variants, can serve as the conversational engine, all accelerated by Apple's MLX framework. Voice cloning from under 10 seconds of audio is supported natively, a feature typically gated behind commercial subscriptions. The companion desktop app is built with Tauri, React, Tailwind CSS, TypeScript, and Rust, and the v0.1.0 release is available as a macOS DMG tested across M1 Pro, M3, and M4 Pro hardware.
The "Cards and Stories" abstraction at the heart of OpenToys borrows directly from Yoto's card-based content model but opens it to anyone with an <a href="/news/2026-03-15-pycoclaw-brings-full-openclaw-agent-loop-to-esp32-microcontrollers-for-5">ESP32</a>. Users define named AI personas — examples from the project include Math Bear, Cosmo the Monkey, and Coach Carter — that children can interact with conversationally. The ESP32-S3 acts purely as a microphone and speaker endpoint, with all inference running on the paired Mac; this architectural split keeps hardware costs low and the data boundary unambiguous. Firmware is flashed directly from the app, and device onboarding uses a WiFi captive portal, making the setup accessible without specialized embedded development experience.
Released under the MIT license, OpenToys commits to community ownership with no proprietary core. The project's GitHub repository at akdeb/open-toys includes firmware guides for both PlatformIO and Arduino IDE workflows, with documentation hosted at elatoai.com. That the same <a href="/news/2026-03-14-runanywhere-launches-rcli-on-device-voice-ai-with-proprietary-metalrt-inference">voice-interactive, multilingual agent experience</a> that runs on a commercial toy subscription can now run on a $10 microcontroller and a MacBook — with no data leaving the room — tells you something concrete about where edge inference is headed.