The r1o Guide

Sign in, download r1o, drag it to Applications, and open it. That is the whole install: the agent, asmi, and model management come bundled. No Homebrew, pip, or Docker.

On first launch, r1o scans your ~/Models directory and any Hugging Face cache automatically. If you already have MLX models downloaded, they show up instantly. Pick one from the Models tab, hit Deploy, and start chatting.

Tailscale creates an encrypted WireGuard mesh between your devices. Install it on your Mac and sign in — that is all. No port forwarding, no firewall rules, no cloud relay. Every device on your Tailscale network can reach your Mac directly.

With Tailscale on each device, your Macs reach each other on a private network, wherever they are. The iPhone and iPad app is coming soon: it will find your Mac over the same network, and the model keeps running on the Mac.

Open r1o on a second Mac. If it is on the same Tailscale network, it joins automatically. The dashboard shows both nodes in the topology view — chip model, RAM, running models, everything visible from one screen.

Transfer models between nodes directly from the dashboard. Over Thunderbolt 5, transfers run at 5.5 GB/s. Over LAN, it falls back to rsync — slower, but it just works. Deploy a model to any node with one click.

Connect your Macs with Thunderbolt 5 cables. From macOS 26.2, Macs can use RDMA over them: direct memory access at Thunderbolt 5 speeds. That is how the cluster moves model weights between Macs.

The topology view shows every active RDMA link. On our own four-Mac cluster, model transfers ran at 5.5 GB/s over RDMA, against about 100 MB/s on gigabit Ethernet: a 40 GB model in about 7 seconds instead of about 7 minutes.

Some models are too big for a single Mac. GLM-5.2 at 743 billion parameters needs over 400 GB of RAM just for the weights. No single machine can run it. JACCL splits the model across RDMA-connected nodes — you deploy, it figures out the sharding.

The Models tab in Mission Control is more than a file browser. Search by natural language — "fast coding model under 50GB" — and get ranked results with RAM fit estimation for every node in your cluster.

Download models directly from Hugging Face, delete old ones, transfer between nodes. The model registry tracks everything: size, quantization, architecture, community benchmarks, which nodes have a copy.

The r1o Agent ships inside r1o, so there is nothing else to install. It runs on your Mac, speaks MCP, and can use every tool in the app: model management, cluster operations, and your knowledge base. Its model can be local, or a provider you connect.

Teach it your workflows with custom skills: markdown files that describe a procedure, which the agent reads and follows. With a local model, your conversations stay on your network.