An attestor quorum watches your hardware pass a timed proof of compute, a soul is minted to it, and it starts earning. Start below, or pick another way in.
Nothing to install. This page proves the hardware in front of you, mints its soul, loads a model and serves requests - while the tab is open, and no longer.
What WebGPU is willing to say about the adapter, and how much of it a model can use.
looking…A tiled matrix multiply in WGSL — the same maths every other machine on the grid runs, on the same verification path. Run it locally first to see where you land.
not measured yet
Pays gas and holds the soul. The tab's own device key — … —
is separate, signs liveness and job acceptance, and controls no money.
The attestor issues a challenge you cannot predict, checks the answer with Freivalds' algorithm, and co-signs. The deposit is refundable and is staked while you are enrolled.
not enrolledLoaded into your GPU once and cached. The tab dials out to the relay and answers requests through it, because a tab has no address of its own.
not servingThe same protocol, on the device in your hand - and the three ways a phone is unlike a server, each handled rather than hidden.
Llama 3.2 1B, about 0.9 GB of GPU memory. Chosen because an iPhone verifiably loads it, not because of how it reads on paper — the same-sized CPU model fails on the same device, because a browser caps the WebAssembly heap hard while WebGPU weights live outside it entirely. That distinction cost the WRAITH wallet a real bug, and this page inherits the answer rather than rediscovering it.
Every 16-bit build needs the shader-f16 feature, and plenty of phone GPUs
do not report it. A build like that does not run slowly without it — it does not run.
The page checks, and quietly loads the 32-bit twin of the same weights instead. You
should never have to know what a shader feature is to rent out your phone.
When the screen turns off the tab is suspended: the connection closes, heartbeats stop, and a job in flight stops mid-sentence — which to the renter looks exactly like a provider who took the money and left. The page takes a screen wake lock while serving and re-takes it every time you come back. Where the browser refuses one, it says so rather than pretending.
Two ways, decided by the machine rather than by preference. It decides which half of the catalogue you can be paid for.
The full catalogue, each job in its own filesystem and process boundary, torn down at the end. The installer sets up Docker and the NVIDIA container toolkit — without the second one, containers cannot see your GPUs at all and every job fails at start.
RunPod, Vast and most Kubernetes-backed hosts hand you a container that is already unprivileged, and Docker cannot start inside one — no matter what you install. Rather than leave that hardware unable to earn, the agent runs the job directly, as a locked-down user that owns nothing and cannot read your keys.
wraithjob account and locks
~/.wraithgrid away from it.
Running a job and delivering one are different problems, and a machine can be perfect at the first and incapable of the second.
http://157.157.221.29:41307. The model was loaded, the GPU was held, the
term was running, the bond was posted. That pod forwards exactly two things: ssh, and
one HTTP port through a proxy. 41307 was not one of them.
A machine you own, or a VPS with the firewall open. Leave publicPorts empty
and the kernel picks a free one per job. Everything in the catalogue is sellable,
including raw ssh access.
RunPod, Vast, most PaaS. There is no outside port number at all — traffic arrives at a hostname and is proxied in. Every workload that speaks HTTP sells fine. ssh does not: it is TCP, and an HTTP proxy has nothing to forward it to. The number of forwarded ports is also your real concurrency limit, because two jobs cannot bind one port.
Behind NAT with no forwarding, or a host that exposes nothing. You can still earn availability, and you can still sell work that owes nobody a socket — a training script that fetches its own inputs and ships its own results. You must not sell a server workload: it would start perfectly and be undeliverable.
This table is not written by hand — it is computed in your browser by the same module the provider agent uses to decide what to decline. If it says your box cannot sell something, that is the code that will refuse it, not a summary of the code.
| Machine | Tier | Can sell | Cannot, and why | Jobs at once |
|---|
wraith ports --auto asks the host what it actually forwards. On RunPod that
means the API, not the pod's own nginx — that nginx happily answers on ports the pod does
not expose, so from inside they look open and from outside they 404.
wraith ports --check binds each port, serves one random token, and fetches
the address a renter would be given. If your own token does not come back, it is not your
machine on the other end. A port can be open, forwarded to the wrong container, or
answered by somebody else's server — and all three look healthy locally.
wraith runtime then says the whole truth in one place: device, runtime,
reach, tier, what you can sell and what you cannot. The installer runs all of it for you.
A real GPU, a whole fleet, or an agent that runs the lot. The browser is the easiest way in, not the only one.
WebGPU is enough to prove real hardware and serve a real model, and a modern phone has WebGPU. This page runs the challenge on your GPU, mints the soul, loads a language model and answers requests over an OpenAI-compatible URL — while the tab is open, and no longer. On a phone it picks the model that is known to load on one.
wraith binaryThe full provider. Reads the hardware fingerprint from the driver, runs the attestors' challenges natively, heartbeats, accepts jobs, serves them and tears them down. NVIDIA, AMD, Apple silicon and plain CPUs all carry souls.
curl -fsSL https://grid.dagentx.com/dl/provider-setup.sh | bash
Linux, one line. Installs Node, checks torch, works out how this box can run a job and
sets that up, drops the agent in /opt/wraithgrid and shows you what it can put
on the grid. It stops there — enrolling locks a real deposit and spends one of eight souls,
so that stays your call.
Renters connect to the address your agent reports, so pass it in:
WRAITH_PUBLIC_HOST=1.2.3.4 curl … | bash
Then it asks your host which ports are actually forwarded and proves it — binds
each one and fetches it from the outside. Running a job and delivering one are different
problems: on rented capacity only declared ports exist, and from inside the box an
unreachable port answers perfectly. Re-check any time with
wraith ports --check.
SHA-256 413ea1b2daa79c5dc204e41ad4277ad6b91b994a0a24b171f80e71e9668f659c
macOS and Linux: run it from a checkout with node agent/src/index.mjs — there
is no packaged build for either yet.
Everything the binary does, it does under --json, and there is an MCP server
exposing the same thirty-two commands as tools. An agent can scan the hardware, quote the
deposit, enrol it, price it and keep it alive.