🦙 Become a provider

Your GPU is idle most of the day.

Point OpenHydra at a model you already run, leave it on, and serve the herd while you’re not using it. In return you earn priority for when you need a model bigger than your own, and reputation that keeps the whole network honest. Reciprocity today, paid frontier tiers on the way.

If you’ve got a computer that mostly sits idle, a spare gaming PC, or a Mac that’s always on, you can let it run AI for other people when you’re not using it. In return, you jump the queue for the network’s bigger models when you need them, and you build a good reputation that keeps everyone honest. Free help both ways today, with paid options for big models on the way.

Download & start sharing Join the Discord

Already running Ollama, LM Studio, vLLM, llama.cpp or Exo? You’re most of the way there, OpenHydra just connects what you already have to the network.

Why share

Give tokens, get priority.

OpenHydra runs on reciprocity, not a wallet. What you contribute while idle is what you can draw on when you need more than your machine can hold.

⚡

Earn priority

Give-to-get: serve tokens while you’re idle and bank priority for when you ask for a bigger model. Contributors jump the queue; freeloaders wait.

Help others by running AI while your computer is idle, and you get bumped to the front of the line when you need a bigger model yourself. People who pitch in get served first; people who only take have to wait.

🏅

Build reputation

Every answer you serve is spot-checked and cross-verified. Honest, reliable providers build standing and get routed more work, your reputation is your bond.

Every answer your computer gives is quietly double-checked. Be reliable and honest, and you build a good reputation, so the network trusts you with more. Your good name is what earns you priority.

🌱

Grow the alternative

Every machine that joins makes the herd less dependent on hyperscalers. You’re not renting out a GPU to a middleman, you’re a peer in an open network.

Every computer that joins makes this network a little less reliant on the giant tech companies. You’re not renting your computer to a middleman, you’re an equal part of an open, people-powered network.

What makes a good provider

If it can load a model, it can feed the herd.

You don’t need a data center. The best providers are ordinary machines that are already on.

🍏

Apple Silicon (M-series Macs & Mac Minis)

Quiet, efficient, and unified-memory generous, a Mac Mini is a superb little inference node. Point OpenHydra at the model you already run in Ollama or LM Studio and serve the herd while you sleep.

Modern Macs (and the little Mac Mini) are quiet, power-sipping, and surprisingly good at running AI. Point OpenHydra at a model you already run and let it help the network while you sleep.

🌐

Home servers & always-on rigs

Got a box that’s on 24/7 anyway? That’s the ideal provider, steady uptime builds reputation fastest. OpenHydra punches through CGNAT, so no port-forwarding or VPN to babysit.

Got a computer that’s on around the clock anyway? That’s the best kind of helper, steady, always-on machines build the strongest reputation. OpenHydra sorts out the tricky network connections for you, so there’s nothing to set up on your router.

How to start

Four steps. One toggle.

If you already run a local model, you can be serving the herd in a couple of minutes.

Install the app

Grab OpenHydra for macOS, Windows or Linux. The macOS build is signed & notarized, so it just opens.

Get OpenHydra for Mac, Windows or Linux, just like any other app. On Mac it opens straight away, no warnings.

Point it at your engine

Already run Ollama, LM Studio, vLLM, llama.cpp or Exo? OpenHydra detects it and wraps it. No new runtime, no re-hosting models, and if you don’t have an engine yet, the app can install one for you.

Already run an app like Ollama or LM Studio? OpenHydra finds it and uses it, nothing new to install or learn. Don’t have one yet? The app can set one up for you with a click.

Choose what to serve, then flip Share

Pick which models you’re happy to offer the herd and turn on sharing. Routing, encryption, NAT traversal and verification are all handled for you, nothing to configure.

Pick which models you’re happy to let others use, and switch on sharing. All the tricky parts, keeping it private, connecting safely, checking answers, are handled for you. Nothing to configure.

Leave it on, earn while you idle

That’s it. Serve tokens in the background, build reputation, and bank priority for when you need a model bigger than your own. Stop any time with the same toggle.

That’s it. Let it run in the background, build up your reputation, and bank priority for when you need a bigger model. Stop any time with the same switch.

Is it safe?

You’re sharing inference, not your machine.

Sharing exposes one thing: the ability to run the models you chose. It does not open a door to your files or your network.

🔒

Encrypted end-to-end. Every connection is authenticated with Ed25519 peer identities and encrypted in transit via libp2p Noise and QUIC. Even over a relay, the relay just forwards bytes, it never sees the prompt.

🔒

Everything is encrypted. Messages are scrambled so only the two computers involved can read them. Even when a request is passed along by another computer, that computer sees only jumbled data, never the actual question.

🎯

Scoped to your engine’s API. Peers can only ask the models you opted in to serve for a completion. They can’t reach your filesystem, your other apps, or the rest of your LAN.

🎯

Only the models you chose are shared. Others can ask your computer to run the AI models you opted in to, and nothing else. They can’t see your files, your other apps, or anything else on your network.

👤

No account, no personal data. You join as a cryptographic peer id, not a sign-up. There’s no profile to leak.

👤

No account, no personal details. You join with a random ID, not a sign-up, there’s no profile or email that could ever leak.

🛑

You’re always in control. Choose which models to serve, cap how much you contribute, and stop sharing instantly, one toggle, no restart.

🛑

You’re always in charge. Choose what to share, cap how much you give, and stop instantly, one switch, no restart.

⚖️

An acceptable-use floor. The network enforces baseline usage policies so you’re not serving abuse. Reputation and verification keep bad actors, on both sides, out.

⚖️

Ground rules keep it clean. The network has basic usage rules so your computer isn’t used for anything abusive, and its reputation system keeps bad actors, on both sides, out.

Questions

The honest FAQ.

What does it cost me?

Electricity, and some bandwidth. You’re running a model you can already run, sharing just sends the answers to a peer instead of only to you. You set the ceiling on how much you contribute.

A bit of electricity and internet data, that’s all. You’re running a model you could run anyway; sharing just sends the answer to someone else instead of only to you. And you decide the most you’re willing to give.

Do I get paid?

Today the network runs on reciprocity: you earn priority and reputation, not cash. Paid frontier tiers, where you’re compensated for serving bigger models, are on the roadmap; we’d rather ship them properly than promise specifics early.

Not in cash yet. For now you earn priority and a good reputation on the network. Paid options, where you’re actually compensated for running bigger models, are on the roadmap; we’d rather build them properly than over-promise.

Will it get in my way while I’m using my machine?

Sharing is meant to run in the background on idle capacity. You control how much you offer, and you can pause with a single toggle whenever you want the whole machine to yourself.

No. Sharing is designed to use only your spare power, in the background. You set how much to give, and you can pause it with one switch whenever you want your computer all to yourself.

Do I need a static IP or to forward ports?

No. OpenHydra uses libp2p with DCUtR hole-punching to connect peers directly through CGNAT, and falls back to an encrypted relay when it can’t. There’s nothing to configure on your router.

No. OpenHydra handles all the fiddly network stuff for you, it connects computers directly through home routers automatically, and quietly falls back to a safe relay if it can’t. Nothing to change on your router.

Which engines can I serve with?

Ollama, vLLM, LM Studio, llama.cpp and Exo are supported today via thin adapters, keep the exact stack you already have. See the docs for setup per engine.

Ollama, LM Studio, vLLM, llama.cpp and Exo all work today, keep whatever you already use. The docs have setup notes for each.

Can peers see my prompts or data?

Peers only send you their prompts to run on the models you chose to serve; they get back only the completion. They can’t see your files, your other requests, or anything else on your machine.

No. Other people only send their questions for your computer to answer, and they get back only the answer. They can’t see your files, your own chats, or anything else on your computer.