PAIR routes subagent work to idle Macs and PCs

Nvidia on September 3 released a beta of the Personal AI Router (PAIR), open source software that sends an agent’s subagent requests to idle Macs and PCs on the same home network. In Nvidia’s own example, two PCs with RTX 5090 GPUs running Qwen3.6 35B A3B sped up work with five subagents by about 1.6x.

PAIR is built for agents like NemoClaw, OpenClaw and Hermes, The New Stack reports. A lead agent splits a task across subagents, and their model requests can then run on whichever machines in the house are sitting idle. That lets the agent run more subagents in parallel.

Nvidia calls PAIR a “virtual inference router” and is explicit that it is not a new inference engine. The models still run in the Ollama or LM Studio install already on each machine. PAIR proxies Ollama-compatible and OpenAI-compatible endpoints, the project’s GitHub repository says, so the agent keeps one configured local address and the router decides where each job runs. Machines with PAIR installed find each other on the local network over mDNS, the zero-config discovery protocol home printers use, and PAIR checks whether each one can take a given request.

Nvidia describes the flow this way: “PAIR receives the request through its proxy, identifies its engine and model requirements, and selects one eligible node. That node executes the request from start to finish and sends the response back through PAIR. The agent continues to see one connection while PAIR handles placement behind it.”

That description sets the main limit. PAIR doesn’t merge GPUs, doesn’t pool VRAM into a single accelerator and can’t split one inference request across machines. Every model you want to route has to fit on a single node, and the speedup comes from running more requests at the same time. A single long chat with one big model gets nothing from a second PC.

Routing is also strict. A request goes only to a machine that has the right engine enabled and the exact requested model available. Each machine can host different models, and putting the same model on several machines gives the router more places to send concurrent requests.

Idle really does mean idle. PAIR tracks which computers are available, and when someone sits back down at a machine to work or game and takes the GPU back, PAIR stops the local inference engine there.

Which machines qualify

PAIR runs on Windows, macOS and Linux. On the Nvidia side, GeForce RTX 20 series GPUs are the baseline. Apple machines need M4 silicon or newer. Nvidia’s DGX Spark is supported, the same box that in August was the $4,800 entry ticket for Perplexity’s local agent, and RTX Spark PCs and laptops join once they ship later this year. Mac support matters here because Macs have become a common way to run local models and agents like OpenClaw, even though they have no Nvidia GPU at all.

The project is licensed under Apache 2.0 and ships signed installers for Windows, Debian and macOS, with a desktop GUI and a terminal interface. Source builds are documented too, and Windows on Arm is experimental.

The 1.6x figure comes with its conditions: two PCs with RTX 5090 GPUs, each with 32 GB of memory, running Qwen3.6 35B A3B with five subagents. The New Stack puts the 5090 at around $5,000 right now, against a launch MSRP of $2,000, which makes the demo rig one that few home users could copy. The outlet also notes that few people have several 5090s at home, so it remains to be seen how PAIR does with a more typical mix, such as a Mac Studio, a few Mac minis and a gaming PC.

On privacy, the repository says prompts and responses are meant to stay on the local network as long as every configured client, model source, engine and node is local. You still need an internet connection to download models.

Getting it running

Nvidia’s getting-started docs lay out the setup. Install PAIR on every machine you want in the pool. Let it discover and pair those systems. Then make sure each one has Ollama or LM Studio installed with the models already downloaded. The repository also lists llama.cpp as an engine, and the docs say PAIR can detect existing Ollama and LM Studio installs.

You don’t have to prepare each machine by hand, either: PAIR can install Ollama, LM Studio or llama.cpp on paired machines and start the model downloads there, so the same model ends up on every node that should share the load.