Your GPU is idle most of the day.
Point it at the gateway.
Run the open-source Edge agent on a machine with a spare GPU. Your machine connects out to EveryAPI and answers marketplace requests using a copy of Ollama running locally — no port to open, no public address, no domain, no firewall changes. Approved sellers add a node from the dashboard and get a one-time setup code.
The tunnel goes outward, which is the whole point.
Home and office machines almost never have a public address, and we did not want to ask you for one. Instead your machine opens a connection to us and holds it open, the same way a chat app stays connected.
- 01BuyerSends an ordinary request, in whichever format their tools already speak.
- 02GatewayTidies up the request and picks a node that has the model they asked for.
- 03Your agentPicks up the job on the connection it opened, and checks it is one of the few request types it is allowed to pass on.
- 04Local OllamaRuns on your GPU. The answer streams back the same way it came.
The arrow pointing at your machine is the connection your machine made. We cannot reach into your network — the node only needs to be able to reach api.everyapi.ai, the same as any website.
Your GPU sends words out as it produces them, so the buyer sees the answer appear rather than waiting for the whole thing. If a connection drops midway, that request is counted as failed — not billed as if it finished.
Follow-up messages in the same conversation come back to the same node, which keeps it fast because your machine still has the earlier context loaded. Nodes joining or leaving only affects conversations that were on those nodes.
You can point any number of machines at a single listing. It offers everything your machines have between them, so they don't all need the same models installed.
Your node checks in regularly. If it stops — power cut, reboot, connection lost — we notice, mark it offline, and stop sending it work until it comes back. Nobody has to do anything.
One command on the machine with the GPU.
Add the node in the dashboard to get a one-time setup code, then run the installer on the machine. It works out what hardware you have, sets everything up to match, and starts it.
Create the node
Approved sellers add an edge node in the dashboard and get a setup code, shown once. We only ever keep a scrambled version of it, and we delete even that once your machine has connected.
Run the installer
It checks how much memory your GPU has and how much disk is free, picks a model that comfortably fits, downloads it, and actually runs a test question through it before telling you it worked. Add --model to choose your own.
Watch it come online
Your machine creates its own private key, proves who it is, and tells us which models it has. Your listing stays switched off until a machine has connected with at least one model that can be sold.
# Values come from the dashboard when you create the nodecurl -fsSL https://dl.everyapi.ai/edge/install.sh | bash -s -- \ --node-id 42 --token edgert_... # Pick the model yourself instead of letting it choosecurl -fsSL https://dl.everyapi.ai/edge/install.sh | bash -s -- \ --node-id 42 --token edgert_... --model qwen2.5:7bA few things the installer does so a first attempt can't leave you stuck: it won't run inside a folder you keep code in, so your setup code can't end up somewhere you'd share by accident; it keeps that code until we've actually confirmed the connection, so a failed first try can be retried instead of leaving a dead node; and running it again later is safe.
Three GPU paths, auto-detected
The installer works out what you have — NVIDIA, then AMD, then Apple Silicon, then CPU — and sets itself up accordingly. If something is missing it tells you and stops; it never installs things on your machine without asking.
The most common setup and the default. You'll need a reasonably current graphics driver and NVIDIA's container toolkit installed first.
Works with AMD cards, as long as you've installed AMD's ROCm on the machine beforehand.
Macs can't share the graphics chip with Docker, so on a Mac we run the model directly on the machine and only the agent runs in a container.
It will run and answer using the processor instead of a graphics card. Fine for trying it out, and genuinely useful for one kind of work (turning text into search-friendly numbers), but far too slow for chat to be worth selling.
You'll need Docker installed, with a current version of Docker Compose — both the installer and the CLI check and give you a link if it's missing, rather than installing anything themselves. The edge agent runs on Linux and macOS.
Everything after setup is one command
One command covers everything a running node needs, and the dashboard shows the same thing. Each node keeps to its own space on the machine, so you can run several side by side.
everyapi edge statusShows what's happening on your machine next to what we see — online or offline, when it last checked in, and which models are installed.
everyapi edge modelsAdd, remove, and list the models on a node. Works without a terminal attached, so you can script it.
everyapi edge pauseStops new work coming in, and stays that way through restarts — nothing quietly puts it back to work. If you run several machines, pausing one leaves the others going.
everyapi edge logs -fWatch what your node is doing, live. The dashboard shows the same thing per node if you'd rather not use a terminal.
everyapi edge updateGets the latest version and restarts. Nothing updates on its own — the software on your machine only changes when you run this.
everyapi edge removeShuts everything down, clears out the files it created, and removes the node unless you'd rather keep it listed. It asks before doing any of that.
So here is exactly what it can and cannot do.
None of this is a promise we're asking you to take on trust — each one is how the software is built, and you can read the code yourself.
On first run the agent creates a private key and locks it down so only you can read it. If that file ever becomes readable by others, the agent refuses to start rather than carry on. We only ever see the public half.
The agent will only forward the small set of requests a model needs to answer, and that limit is built into the agent itself rather than something we configure. Even if our side were compromised, it could not make your machine do something else.
There is nowhere in the agent to put it — the record it keeps has no space for the question or the answer. It notes which model was used, how long it took, how much was processed, and a meaningless label for the customer. No questions, no answers, no keys, no email addresses.
It's shown once, we keep only a scrambled version, and we delete even that after your machine first connects — so it can't be reused, by anyone. After that your machine identifies itself with its own key each time.
New versions only arrive when you ask for them. And if you remove a node from the dashboard, the agent is told to stop for good — it shuts down with a clear reason instead of retrying forever.
You're credited only after a request has actually been paid for, and each request counts once. If your machine was offline when that happened, it picks up what it missed next time it connects. Earnings go to your seller balance, and moving them to your main balance is a separate step.
The agent is open source under Apache 2.0, and all of it is at github.com/everyapi-ai/everyapi-edge — including everything described above.
Questions worth asking first
Anything else, write to support@everyapi.ai — a real person answers.
No. Edge nodes are part of the marketplace, and seller access is reviewed. You apply first and we look at it — things like how long you've had the account and how you've used it, which the operator can adjust. Only after that can you add a node.
Spare GPU cycles,
pointed somewhere useful.
Apply for seller access, register a node, and run one command on the machine. Seller access is granted after we review your application.