Why Basis

The agent economy runs on inference. Today that inference is expensive, opaque, and hard to compose. Basis exists to make inference cheap to call, easy to integrate, accountable per job, and payable on a common rail — without asking anyone to learn a new SDK or trust an unaudited invoice.

Start with the Overview

This page summarizes the rationale; the Overview is the canonical entry point — what Basis is, who uses it, what you pay, what workers earn, what receipts record, and where Bankr fits.

The case in five lines

Agents need composable inference
At agent volume, the marginal cost of a token matters — and inference should wire into a larger workflow over plain HTTP, with each call accounted for to the last base unit.
Developers want OpenAI compatibility
Adopting a provider should be a base URL change, not a rewrite. Basis implements the chat-completions contract, and returns a structured runtime_pending in the same error shape when no backend is configured.
GPU owners need structure
Contributing idle capacity should be a defined role: register with an EVM reward address, advertise the models you serve, heartbeat, take routed jobs, and earn $BASIS for completed work.
Receipts give provenance
A canonical-JSON, SHA-256 receipt records the accounting of every settled job — model, server-counted tokens, charge, and reward (never the prompt or output) — and a job can never be settled twice.
Base settlement is a common rail
Worker rewards settle idempotently on Base, off the inference hot path: a batch settles once, and a failed batch stays recoverable.