Your GPU. One link. The world. Watch our 30-sec YouTube reel:

Share your GPU with anyone, anywhere — no accounts, no cloud, no listings, no middleman.

RunSnack turns any machine with Docker into a one-link terminal you can hand to a specific person, right now. Not a marketplace, not a rental platform — a direct connection between two people, peer to peer.

🔒 Security model & threat model — Quick tip: power off for ~10 seconds between sessions (with different people only), so the GPU's memory is cleared.

---

Why RunSnack exists

Every GPU marketplace — Vast.ai, SaladCloud, RunPod, Akash — solves the same problem: connect a stranger to some GPU, through an account, a listing, and a platform sitting in the middle taking a cut.

That's not the problem RunSnack solves.

RunSnack gets someone specific onto your hardware, right now:

• A student who needs the lab's GPU for twenty minutes.
• A maintainer who needs to reproduce a bug on your Jetson.
• A collaborator who needs the robot in the next building.
• A friend who needs a CUDA box and you happen to have a spare one idling.
• You're simply too thrifty :) and want to save by renting at a lower cost than mainstream cloud providers.
• You want to earn extra money by renting out your unused machines that were just gathering dust.

No account for them to create. No listing to publish. No marketplace to wait for.

How it works

1. Run the installer on the machine you're sharing (Windows, Linux, or macOS — one script, no config).
2. Answer three questions: CPU cores, RAM, an optional access schedule.
3. Get a link. Send it to whoever you're sharing with. They paste it in a browser and get a real terminal, inside a sandboxed Docker container, on your machine — nothing installed on their end.

That's the whole flow. First run pulls a ~1.4 GB image (a few minutes); every session after that starts in seconds.

# Linux / macOS
chmod +x snackup.sh && ./snackup.sh

# Windows (PowerShell)
powershell -ExecutionPolicy Bypass -File snackup.ps1

What's in this repository — and what isn't

This repo holds the two installer scripts — snackup.sh and snackup.ps1 — byte-for-byte the same ones served from runsnack.com/download. Read every line before you run them; that's the point of putting them here.

The scripts themselves are deliberately thin. What they do, in full:

• Check that Docker is installed and running.
• Ask you three questions (CPU, RAM, schedule).
• Pull gunzfanatic/runsnack-agent from Docker Hub and start it with a hardened, locked-down set of docker run flags (read-only root, all capabilities dropped, non-root user — you can read those flags directly in the script).
• Print the link the container hands back.

Everything past that point — establishing the connection, moving the terminal session across it, keeping it alive, tearing it down cleanly — happens inside the container image, which is closed source and not part of this repository. That's a deliberate line: the scripts that touch your shell with elevated intent are fully readable here; the engineering that makes the connection itself work isn't published.

What makes RunSnack different

• Zero network required. Works the moment you install it, with a network of exactly one — you and whoever you send the link to.
• No account, ever. Not for the host, not for the person connecting.
• Nobody takes a cut. RunSnack charges no fees. If money changes hands, that's a private arrangement between you and the other person.
• Seconds to first shell, not minutes of billing setup.
• Your hardware, your custody. The host's machine is never handed to a platform — it stays exactly where it is, sandboxed behind a read-only, capability-dropped, non-root Docker container.
• Runs on anything with Docker — a Jetson Nano, an old gaming rig, a datacenter card, even CPU-only. Not limited to whatever a marketplace happens to stock.
• No crypto, no wallet, no token.

RunSnack vs. distributed GPU marketplaces

	RunSnack	Vast.ai / TensorDock	SaladCloud	RunPod Community	Akash / io.net / Nosana	CloudJetson
Core model	Direct link between two people	Brokered marketplace	Aggregated compute pool	Brokered marketplace	Token-settled permissionless market	Owned fleet, hourly rental
Works with zero network	✅ Yes — share with anyone you know	❌ No — needs listings	❌ No	❌ No	❌ No	N/A
Private / known-party sharing	✅ Yes — primary use case	❌ Not supported	❌ No	❌ No	❌ No	❌ No
Account / signup	✅ None	❌ Required	❌ Required	❌ Required	⚠️ Wallet required	❌ Required
Platform takes a cut	✅ No	❌ Yes	❌ Yes	❌ Yes	⚠️ Protocol fees	❌ Yes (it's the seller)
Time to first shell	✅ Seconds — one script	❌ Minutes + billing setup	⚠️ Minutes	⚠️ Minutes	❌ Minutes + wallet	⚠️ Minutes
Access method	✅ Browser terminal, link only	SSH / API / Jupyter	Container job submission	SSH / API	Container deployment	Remote access to their boxes
Who holds custody	✅ Nobody — host keeps their machine	Platform brokers	Platform orchestrates	Platform brokers	Protocol	Provider
Edge & SBC devices (Jetson, Pi)	✅ Yes — whatever the host owns	⚠️ Rare / none	❌ No	❌ No	❌ No	⚠️ Yes — Orin & Thor only
Hardware breadth	✅ Anything that runs Docker	Consumer + datacenter	Consumer GPUs at scale	Consumer + datacenter	Datacenter-leaning	Jetson only
Crypto / token required	✅ No	No	No	No	❌ Yes	No
Reputation system	Community channel (Discord)	Built-in	Built-in	Built-in	On-chain	N/A

Competitor features and pricing change frequently — verify independently before relying on this table. Compiled August 2026.

Safety notes — read before sharing

• Sessions run inside a sandboxed Docker container: read-only root filesystem, all Linux capabilities dropped, no-new-privileges, non-root user. Whoever you share a link with gets a terminal inside that container, not your host machine.
• Docker is a sandbox, not a hypervisor. Don't share with anyone you wouldn't hand a real (if contained) shell to, and don't run this on a machine with sensitive data you're not willing to isolate behind that boundary.
• The GPU is the weaker boundary. RunSnack passes the GPU through whole — no MIG, no vGPU, no partitioning. VRAM isn't reliably zeroed between contexts, so data can leak across sessions with no container escape involved. This isn't vendor-specific and it isn't a config we skipped: consumer GPUs have no partitioning to configure.
• RunSnack doesn't run, monitor, or moderate sessions. There's no escrow, no payments, no SLA — if money changes hands, that's between you and the other person.
• Read SECURITY.md before sharing with anyone. It's the full threat model: what the boundary does, what it doesn't, and how to reduce your exposure.

Links

• Website: runsnack.com
• Docker image: hub.docker.com/r/gunzfanatic/runsnack-agent
• Discord (find/offer GPU time, support): discord.gg/jaKt2HANNP
• Author: Manuel Herrador, PhD - github.com/PacifAIst

License

The installer scripts in this repository are released under the Apache License 2.0. The RunSnack agent binary and its Docker image remain closed-source and are distributed separately via Docker Hub.

Made with ❤️ for the Local AI Community by PacifAIst