Nvidia has released open-source software that turns unused home computers into a distributed network for running artificial intelligence agents, according to a report published Thursday by The New Stack. The Nvidia Personal AI Router, or PAIR, allows agents like NemoClaw and OpenClaw to delegate tasks to subagents that run simultaneously across multiple idle Macs and PCs on a home network. The technology is designed to accelerate agentic workflows by distributing parallel requests to machines that would otherwise sit dormant, the report explains.
PAIR functions as a virtual inference router that discovers compatible systems on a local network using mDNS and directs model requests to available machines, the report details. The software works with Windows, macOS, and Linux machines equipped with Nvidia GeForce RTX 20 series GPUs or newer, Macs with M4 silicon or newer, or Nvidia DGX Spark systems. Testing with two PCs running RTX 5090 GPUs with 32 GB of RAM and the Qwen3.6 35B A3B model showed work with five subagents sped up by roughly 1.6 times. The router doesn't merge GPUs or pool video memory into one accelerator, and it can't divide a single inference request across multiple machines. Instead, PAIR uses existing Ollama or LM Studio installations on each machine to execute complete requests from start to finish on one selected node.
Nvidia describes the workflow by explaining that "agents can send a request through the familiar local interface it expects," while PAIR receives the request through its proxy, determines engine and model requirements, and picks one eligible node. The system monitors which computers are available and stops the local inference engine when a user returns to work or gaming on a given machine and reclaims the GPU. PAIR will only route a request to a machine with the required engine enabled and the exact requested model available, meaning installing the same model on several machines gives the router more options for distributing concurrent requests.
The software aims to make local AI agents more practical by leveraging hardware that already exists in households, particularly as Nvidia continues its push toward open models and on-device inference following its acquisition of Hugging Face. PAIR is now available as a beta, requiring users to install it on every machine they want included in the network, have it discover and pair those systems, and ensure Ollama or LM Studio is installed with models downloaded. The router can also assist with setup by installing Ollama or LM Studio and initiating model downloads on paired machines. While Nvidia's testing used high-end RTX 5090 cards that currently sell for around $5,000 despite a $2,000 manufacturer's suggested retail price, the report notes that even a mix of a Mac Studio, a few Mac minis, and a gaming PC should accelerate a local agent workflow. The bet on distributed home computing for AI represents a shift toward making advanced agent capabilities accessible without cloud dependence or enterprise infrastructure. For businesses evaluating on-premises AI strategies, the question isn't whether local inference will improve but whether the coordination overhead justifies the performance gains in production environments.