Connect
Rent compute or earn from your own hardware — from a browser tab, a desktop app, or an agent that does it for you. Same protocol underneath, same jobs, same money.
Connect a wallet and rent in a tab. Keys for your job are generated in the browser and never leave it — the endpoint is sealed to a key only you hold.
Rent: pick a workload, pick hours, escrow. You get URLs back.
Earn: the browser cannot read your hardware's fingerprint or drive Docker, so providing needs the desktop app or CLI. The dashboard still shows your cards, earnings and licences.
Open the dashboard
A single executable that does both sides. wraith ui opens a local
control panel that checks your machine and walks you through enrolling.
wraith ui control panel, in your browser wraith enroll prove the card, mint its soul wraith run stay alive, take jobs
Needs Docker, plus whatever your hardware's class needs - the NVIDIA driver,
ROCm, macOS, or on a CPU nothing beyond Python with numpy. The full
list is under "What you can bring", below.
SHA-256 413ea1b2daa79c5dc204e41ad4277ad6b91b994a0a24b171f80e71e9668f659c
Unsigned, so Windows will warn on first run. Verify the hash before trusting it.
On Linux or macOS, one line instead:
curl -fsSL https://grid.dagentx.com/dl/provider-setup.sh | bash
Installs the agent, works out what this box can run, and stops before enrolling — that locks a deposit and spends one of eight souls, so it stays your call. What it does, step by step.
The same binary is a scriptable CLI, and ships a Claude Code skill so an agent can rent a GPU for a task, use it, and settle up — without you in the loop.
wraith workloads wraith browse wraith rent 3 2 ollama model=llama3.1:70b wraith endpoint 12 wraith confirm 12
The skill tells the agent to quote you a price and wait for a yes before spending anything.
Repo (private)What you can bring
Four classes of machine can carry a soul, each probed a different way and each worth a different amount. The difference is priced rather than papered over: an hour of a CPU is not an hour of an H100, and pretending otherwise would only mean renters paying GPU money for CPU work once and never coming back.
Capability and reward
Everything below is computed from the same table the attestor and the market use. WCU is the unit: 0.55 x dense FP16 throughput, 0.30 x memory bandwidth, 0.15 x capacity, scaled by class and normalised so 1.00 is roughly one RTX 4090. Deposit, reward weight and capability tier all follow from it, so none of these are marketing numbers - they are what the contracts will do when you enrol.
| Device | Class | Tier | Memory | WCU | Reward weight | Deposit | Suggested |
|---|
An unrecognised CPU still enrols, on a generic template scaled to the cores and memory it actually has. An unrecognised accelerator does not: deposit, reward weight and the attestor's throughput band all key off the model, and a guess there is a mispriced listing rather than a cosmetic gap. Tell us what you have and it gets added.
Workloads
Every surface reads the same catalogue, so "run Llama 70B" means the same thing in the browser, the desktop app and an agent script. The provider's machine resolves the same entry and only runs images from that list unless its operator opts in.
| Workload | What you get | Runs on | Reach | Needs |
|---|
Anything not on the list can still run through raw — a box and your SSH key
— if the provider has allowed custom images.
What each surface can do
| Capability | Browser | Desktop | CLI / agent |
|---|---|---|---|
| Rent compute | ✓ | ✓ | ✓ |
| Fetch & decrypt endpoints | ✓ | ✓ | ✓ |
| Confirm, dispute, reclaim | ✓ | ✓ | ✓ |
| Buy & sell licences | ✓ | ✓ | ✓ |
| Enrol a machine | — | ✓ | ✓ |
| Prove liveness & take jobs | — | ✓ | ✓ |
| Withdraw earnings | ✓ | ✓ | ✓ |
The two gaps are physics, not policy: a browser tab cannot read a device's UUID and serial from the driver, and cannot drive Docker on your machine. Everything that does not need the hardware itself works everywhere.
