Anywhere

Inspiration

I wanted to make games in Unity, but my computer couldn’t handle it. My GPU was too weak, so the editor lagged, builds took forever, and anything with real lighting or 3D scenes was basically impossible. Buying a better computer wasn’t an option, and when I looked into renting a GPU in the cloud, it was overwhelming. There were hundreds of machine types I didn’t understand, setup guides full of drivers and SSH commands, and the constant worry of forgetting to turn it off and getting a huge bill. I realized a lot of people are stuck in the same spot: students, artists, and beginners who have the ideas but not the hardware. That’s why we built Anywhere. We wanted getting a powerful GPU to be as simple as saying what you want to do.

What it does

Anywhere turns a plain-English request into a real GPU machine. You type something like “set up Blender on a GPU” or “fine-tune a small model on these support tickets” in the web app, or text it over iMessage. An NVIDIA Nemotron agent figures out what you’re trying to do, and a deterministic planner picks the right machine from NVIDIA Brev’s live catalog and estimates the cost, so the AI never makes up hardware specs or prices. You see the price and approve it, then Anywhere starts the machine on Brev, installs everything, and restores your saved project files. You can watch the desktop live in your browser, take control yourself, or hand it back to the agent and let it work for you, and on your phone you get updates and screenshots over iMessage. When you’re done or go idle, it saves your work and shuts the machine down so you stop paying. Anywhere also works as an MCP server with secure sign-in, so AI assistants like Claude can use it as a tool. Spending caps, approval by default, automatic idle shutdown, and cleanup of failed machines keep anyone from getting a surprise bill.

How we built it

NVIDIA Nemotron powers both the main agent, built with LangGraph and typed tools, and the desktop operator. The operator works in a loop: it looks at a screenshot, Nemotron picks one action, the action runs on the desktop, and it checks the result. NVIDIA Brev provides the GPU machines. We control them through the Brev CLI and only offer machine types Brev currently lists. SpacetimeDB is our database: every table is private, each browser only sees its own data through live views, and rules like “spending can only go up” are enforced inside the database itself. Chat, machine status, live costs, logs, and agent steps update instantly in every open tab. Photon powers the iMessage bridge in both directions. The backend is Python with FastAPI. It handles planning, budget rules, a supervisor that starts, stops, resumes, and cleans up machines, and a secure VNC proxy with short-lived tickets, so a GPU machine is never exposed to the internet. The frontend is Next.js with shadcn/ui and noVNC for the live desktop; project files are stored in Cloudflare R2, and everything deploys with Docker Compose and Caddy on one small server.

Challenges we ran into

The hardest problems were about money, because real GPUs cost real money. We found cases where a training job could slip past the budget guard, the cleanup could delete a machine that was still starting up, two quick approvals could both get past the one-machine limit, and a crashed training job could keep billing for hours. We tracked down and fixed each one with tests. Brev’s catalog was another challenge: some machine types appear and disappear with capacity, so we had to check the live list and refuse plans for machines Brev no longer offered. One Nemotron model was listed in NVIDIA’s catalog but returned errors when called, which taught us to track whether real calls succeed instead of trusting the list. Streaming a GPU desktop securely to the browser through SSH tunnels and an authenticated proxy, with clean reconnects, took many rounds. Because everything updates live, we also had to make sure each database table was subscribed once instead of once per component, and that timers and animations only run when something on screen is actually changing.

Accomplishments that we're proud of

We’re proud that the whole flow works from one sentence to a GPU machine, with the AI handling what you want and a deterministic planner handling the hardware and the price. We built spending safety into every layer: caps, approvals, idle shutdown, spending that can never go down, billing that survives restarts, and cleanup on every failure path. The experience is live across devices, with every browser tab updating instantly and the same agent available over iMessage. The agent can actually operate a real desktop, and you can take control and hand it back at any time. We also built a proper MCP server with OAuth, so other AI assistants can use Anywhere, and backed it all with a large test suite of more than 169 backend tests plus database tests, checking our fixes against real services wherever we had access.

What we learned

We learned to let AI decide what the user wants, but not the facts. Keeping hardware choices and prices deterministic made the product much more trustworthy than letting the model guess. We learned that when real money is involved, safety has to be part of the design from the start, and that most of the important work happens in parts users never see, like guards, cleanup, and billing accuracy. We also learned that real-time apps can fail quietly: duplicate subscriptions, forgotten timers, and repeated polling don’t break anything you can see, but they slow everything down. Finally, we learned to test against the real service, because the docs and model catalogs don’t always match what actually works.

What's next for Anywhere

Next, we want to take the Brev desktop and training paths from tested to running on real GPUs every day. Then we want to bring things full circle with support for Unity and other game engines, so someone with a weak laptop can open a powerful game-dev machine with one sentence. We also plan to add more one-click templates like ComfyUI image generation, Jupyter with PyTorch, VS Code, and “fine-tune, then chat with your model.” Hardware-accelerated desktop graphics would make 3D apps feel native. Support for more GPU providers would let Anywhere find the best price across clouds, voice control would let you talk to your machine over Photon, and shared machines and budgets would let whole teams build together.