Skip to main content
Built for Physical AITuya / Embedded AI infrastructure

Let AI
connect to the physical world.

Agentic-Kit is a lightweight, modular device-side SDK for Physical AI.
From low-compute MCUs and Linux SoCs to PCs, mobile apps, and browsers. Connect voice toys, embodied robots, and more through one real-time device-to-cloud integration, adding perception, Agents, memory, and models as needed.

Connect your hardware Explore capabilities
Modular capabilities · Chip agnostic · Cloud and model agnostic · No changes to business logic · Use what you need
~120 KBSDK sizeFits low-compute devices
~60 KBRuntime memoryLeaves room for your logic
30-70%Packet loss toleranceT-RTC network resilience
≤86 msGlobal average latencyReal-time device-to-cloud transport
~1.5 sEnd-to-end voice latencyWith long-term memory and knowledge base

Typical figures. Actual performance varies by chip, feature configuration, network, and product design.

01 / ONE KIT, A COMPLETE FOUNDATION

Open building blocks.
Room for your ideas.

Agentic-Kit is the open device-side entry point to the Tuya Physical AI platform, supporting AI innovators from development and mass production to global growth.

YOUR AI DEVICE

Your AI hardware

  • Chip, BSP, and peripheral drivers
  • Microphone / Camera capture
  • Speaker / Screen / Actuators
  • Your firmware and product logic
AGENTIC-KIT

Real-time device-to-cloud SDK

  • Provisioning, Activation, and authentication
  • Real-time multimodal transport
  • Session events and callbacks
  • Device-side MCP command exchange
TUYA PHYSICAL AI

Tuya Physical AI Platform

  • Speech and vision algorithms
  • Foundation models, Agents, and memory
  • Device control and scene services
  • Global connectivity and ongoing operations

Choose what you need. Know who owns what. Choose platform models, algorithms, memory, and industry services as needed. Connect your own core cloud services through open interfaces.

02 / DESIGNED FOR EMBEDDED CONSTRAINTS

Built for AI.
Grounded in hardware.

Device-side integration is more than connecting. It must stay usable with limited resources, diverse chips, and unstable networks.

/ 01

Keep AI in the cloud and devices lean

Keep capture, playback, protocols, and local control on the device. Offload complex inference to the cloud as needed for a complete AI experience within limited RAM and Flash.

Lightweight devices / Cloud AI
/ 02

Isolate chip differences with the PAL

The PAL abstracts networking, threads, time, and storage. Change chips or operating systems without rewriting your business logic alongside the platform code.

PAL / Port once / Keep your logic
/ 03

Keep interactions flowing on poor networks

Build a real-time link for audio, images, video, text, and events, with handling for streaming, interruptions, reconnection, and network fluctuations.

Streaming / Interruption / Reconnection
/ 04

Let the device decide what runs

AI sends tool requests as MCP commands. Local firmware decides whether to execute, performs the action, and returns the result.

MCP / Local decisions / Execution results
Available platforms and porting paths
ESP32 / ESP-IDF · FreeRTOS · ARM · MIPS · Linux
03 / COMBINE THE AI YOUR PRODUCT NEEDS

One foundation.
Many kinds of AI hardware.

From companion, learning, and productivity devices to embodied robots. Combine conversation, perception, memory, execution, and device ecosystem capabilities around your product.

CapabilitiesExample scenario setup
PerceptionVAD / PVAD / ASRIncluded
VoiceTTS / Character voicesIncluded
Cloud AILLM / Character profilesIncluded
Memory & knowledgePreferences / StoriesOptional
Local toolsStatus light / ExpressionsOptional
Your servicesCharacter IP / ContentCustom

Hardware: microphone, speaker, and status light. Firmware handles capture, playback, and local feedback.

Product experienceScenario flow

Toys that talk, with a personality of their own.

Hear what children say, respond in character, and bring each reply to life with voice and device actions.

"Tell me a story about going to the moon."
  1. 1The device captures speech and uploads it through the SDK.
  2. 2The cloud recognizes speech and creates a story and voice response in character.
  3. 3The device plays the reply while firmware controls the status light.

Platform capabilities. Your product. Agentic-Kit provides connectivity and data exchange. Configure algorithms, memory, and industry services to suit your product.

04 / PERCEPTION FOR REAL DEVICES

Algorithms prove their worth
on real hardware.

Physical AI brings together on-device perception, cloud algorithms, and real hardware. Agentic-Kit lets you deploy and combine algorithms to suit your product.

01 / ON-DEVICE PERCEPTION

Perception built around your hardware.

On-device solutions can provide keyword wake-up and visual preprocessing. Run VAD on the device or in the cloud to suit your product.

KWS · VAD · Visual preprocessing
02 / TUYA CLOUD ALGORITHMS

Cloud algorithms that keep improving the experience.

Send audio, images, and video to cloud algorithms over a real-time link and receive text, speech, or structured results.

Nomo-ASR · Nomo-PVAD · Pet ID
03 / SPECIALIZED MODELS

Specialized algorithms for the real world.

Continuously optimized for real device challenges: far-field audio, competing speakers, individual pet recognition, and complex scenes.

Speech models · Audio models · Vision models
05 / HELLO, PHYSICAL WORLD

Get it running.
Then make it yours.

No chip selection or development board required. Run the text chat example on macOS or Linux to test the device-to-cloud link, then bring the proven interaction to your hardware.

  1. 01
    Test on your computer
    Get the source and run the POSIX example
  2. 02
    Move to your hardware
    Adapt the platform, audio/video, and local tools
  3. 03
    Prepare for production
    Configure the PID, device identity, and product capabilities
Read the integration guide
Terminal · macOS / Linux
# Get the source and submodules
git clone https://github.com/tuya/agentic-kit.git
cd agentic-kit
git submodule update --init --recursive

# Build and run the POSIX text chat example
cmake -S examples/posix -B build-examples
cmake --build build-examples
./build-examples/text_chat_demo
CMake ≥ 3.20 / Python 3 / C and C++ toolchain

Use the desktop example to test the interaction path. For your product, adapt the platform, peripherals, and device identity to the target system.

RTC TCP Client ->Source integration · PAL portingRTC Client ->Prebuilt library · Target platform matching
06 / YOUR AI, YOUR WAY

From ready-made frameworks
to flexible building blocks.

TuyaOS, TuyaOpen, and Agentic-Kit offer different device-side frameworks and integration paths for different project stages and technology stacks.

STANDARDIZED / PRODUCT FOUNDATION

TuyaOS (Wukong)

Tuya's standardized product framework integrates connectivity, security, device models, and mass production support.

  • A framework that fits your project
  • Capabilities scoped to your product
  • Support from development to mass production
Explore TuyaOS (Wukong)
STANDARDIZED / OPEN-SOURCE STACK

TuyaOpen

Tuya's standardized layered framework, with open source code and a broad set of AI and IoT components.

  • A framework that fits your project
  • Capabilities scoped to your product
  • Support from development to mass production
Explore TuyaOpen
MODULAR SDK / FITS YOUR STACK

Agentic-Kit

Keep your existing stack and design choices. Select only what you need for resource-constrained or deeply customized products.

  • A framework that fits your project
  • Capabilities scoped to your product
  • Support from development to mass production
Explore Agentic-Kit

Your next AI device starts here.

From embedded integration to platform capabilities on demand.

Explore Agentic-Kit