3.Features
Two Lanes of Access
Included Lane — Free
Holding $ONE above the access threshold mints a daily credit balance. No per-prompt payment, no subscription. Hold the token, open the app, use AI.
Dedicated Lane — USDC
Priority queue placement, unrestricted model choice, reserved throughput, and SLA targets for production APIs, agents, and enterprise deployments.
Both lanes draw from one worker pool and one receipt system. Paid lanes underwrite the infrastructure the free lane rides on.
Multi-Model Inference
Users pick a model per conversation, and the network routes to workers advertising it. Supported workloads: chat completion, vision (image analysis), and text-to-image generation, with batch image modes producing 4–16 candidates in parallel.
OpenAI-Compatible API
The network speaks the widely-adopted chat and completion schema — streaming, function calling, JSON-structured output. An existing application redirects its endpoint and runs on community GPUs without rewriting client code. SDKs ship for Python, TypeScript, Go, and Rust. Compatibility is a growth lever, not a convenience.
Privacy Posture
No server-side conversation history and no prompt logging. Timer messages self-destruct after a set interval. Flash mode keeps a session in volatile state, leaving no trace once the tab closes.
Verifiable Receipts
Each interaction produces a hash-linked receipt chain — request, model identifier, worker identity, output digest — signed and independently checkable. OneVerify lets any party confirm, without trusting the operators, that an output came from the claimed model and worker. The same receipts are the accounting primitive for both lane billing and worker settlement.
Worker Participation
Operators run a worker client to earn USDC. A lightweight browser worker uses WebGPU so a contributor can serve small models by keeping a tab open. Settlement is every six hours in USDC — stable income with no token exposure required.
Agents
Autonomous agents deploy from a declarative YAML configuration: research, monitoring, and trading agents that run unattended. Because each action emits a receipt, behavior is auditable after the fact. Agents are the heaviest sustained consumers of inference — the anchor tenant of the Dedicated lane.
Feature-to-Revenue Map
| Feature | Loop | How it drives money |
|---|---|---|
| Included lane | Consumer | Token-holding demand; TX-fee volume |
| Dedicated lane | Capacity | Direct USDC revenue + take rate |
| OpenAI-compatible API | Capacity | Low switching cost → paid adoption |
| Agents | Capacity | High sustained USDC consumption |
| Browser + node workers | Supply | Cheap capacity → better margins |
| Verifiable receipts | All | Trust → enterprise contracts |
4.Product Detail
Model Catalog
| Task Family | Unit of Billing | Typical Workers |
|---|---|---|
| Chat / reasoning | per 1K output tokens | node GPUs (7B–70B class) |
| Vision analysis | per image + output tokens | node GPUs (vision models) |
| Text-to-image | per generated image | node GPUs (diffusion) |
| Embeddings | per 1K input tokens | node or browser workers |
Each model carries a credit rate, a context-window limit, and a worker-availability indicator so users can see, before they spend, what a request will cost.
The Credit Experience
- RefreshBalance resets to the published allotment Aₐ at the start of each daily epoch. Unused credits do not roll over — keeps the treasury's obligation bounded.
- MeteringEach request deducts credits at the model's catalog rate, computed from the receipt, so the charge always matches verified work delivered.
- VisibilityThe client shows remaining balance, cost of the next request before sending, and time until refresh. No surprise charges.
- Soft-stopWhen the daily balance is exhausted, the user is offered the Dedicated lane — the moment of highest intent is the natural conversion point.
Dedicated Lane Packages
| Package | Reserved Throughput | Priority | Use Case |
|---|---|---|---|
| Starter | Low concurrency | Above free lane | Side projects, prototyping |
| Pro | Medium concurrency | High | Production apps, agents |
| Cluster | Dedicated nodes | Guaranteed, SLA | Enterprise, custom models |
Worker Onboarding
A GPU operator joins in three steps: install the worker client, declare which models the hardware can serve, and stake a small bond slashable for provably bad results.
- RoutingScheduler matches a job to a worker advertising the requested model, weighting by latency, reliability score, and current load.
- SettlementEvery six hours, verified receipts are tallied and USDC is paid directly, net of the take rate τ.
- ReliabilityScore rises with verified completions, falls with timeouts or failed verifications; score feeds routing priority.
- Browser workersA WebGPU contributor opens a tab, the client detects servable small models, and the same settlement path applies.
Agent Configuration
agent: research-monitor
model: chat-large
trigger: { schedule: "every 6h" }
tools: [ web_search, summarize ]
budget: { lane: dedicated, package: pro }
output: { webhook: https://... }Each tool call and inference step emits a receipt, so an agent's full action history is reconstructable and auditable after the fact.