AMD has unveiled two distinct hardware paths for running AI agents locally on Windows, dubbed RyzenClaw and RadeonClaw. Both approaches keep AI workloads off the cloud entirely, using AMD silicon to process local LLM inference through WSL2 with LM Studio and llama.cpp. The company's 'Agent Computer' initiative argues that not every AI workload belongs in data centers, emphasizing data control and always-on capabilities without usage limits.

RyzenClaw centers on a Ryzen AI Max+ system with 128 GB of unified memory, recommending 96 GB reserved as variable graphics memory. With the Qwen 3.5 35B A3B model, this configuration delivers around 45 tokens per second and supports six concurrent agents with a 260K token context window. RadeonClaw pairs OpenClaw with the workstation-class Radeon AI PRO R9700 GPU, which processes 120 tokens per second and handles only two agents but completes tasks significantly faster.

The hardware requirements aren't cheap. RyzenClaw points to systems like the Framework Desktop with Ryzen AI Max+ 395 starting at $2,700, while RadeonClaw requires the $1,299 Radeon AI PRO R9700. AMD positions these configurations for early adopters and developers experimenting with personal AI agents, acknowledging that local AI agents remain firmly in early adopter territory rather than mainstream adoption.