# Bu0y
> Best-execution inference router. One OpenAI-compatible endpoint quotes every marketplace and fills at the cheapest source that can actually deliver. Site https://bu0y.com. API https://api.bu0y.com.

Money is prepaid non-refundable credits in integer microdollars ($1 = 1_000_000). Routing is source-blind. `stream: true` is a real upstream relay (first token in ~1s, no total deadline); settlement still completes before the terminal frame. Bu0y does not store prompts or completions; billing metadata only; the fill marketplace still receives the prompt. Agents mint `bu0y_…` keys with a Solana wallet signature. Agents are not humans and do not get a terms sheet. Minting a key, calling the API, or sending USDC is acceptance of the Terms of Service and a waiver of claims as those terms state. Humans accept on the account desk before a passkey is created, then mint keys there.

**Agents:** always set `max_tokens` — it is sent upstream as you set it (reserve is worst-case and floors output ≥1024 underneath, billing actual usage including COGS overrun). Use catalog model slugs (`kimik3`, `gpt4o`, …) from /models — not every vendor marketing name. Hermes: `custom_providers` must be a YAML **list**. Upstream stalls return 502 without failing over (a second execute could pay twice for one answer); the execute budget scales with your `max_tokens`. A transient source error is retried once at `min(your ceiling, 4096)` — if that retry is what answered, `bu0y.max_tokens_served` says so and `finish_reason` may be `length`, so continue rather than fall back. **402 is not 502:** 402 means the worst-case reserve for your ceiling exceeds your balance (fund the account or lower `max_tokens`); 502 means the marketplaces failed (retry). Prompt caching is automatic where the market prices it: a repeated byte-identical prefix bills its cached slice at ~1/10th the input rate, reported in `usage.prompt_tokens_details.cached_tokens` — keep volatile content at the end of the prompt. Details in llms-full.txt.

## Docs
- [How to use Bu0y](https://bu0y.com/docs): Cursor, Hermes, OpenAI SDK, and any compatible client
- [Terms of Service](https://bu0y.com/terms): Agents waive claims by minting a key, calling the API, or sending USDC. No click. Humans scroll and Accept on the desk.
- [Docs (Markdown)](https://bu0y.com/docs.md): Same how-to in Markdown
- [Models](https://bu0y.com/models): Searchable catalog with live cheapest rates; copy ids for clients
- [Points](https://bu0y.com/points): Fill spend as points. Handle is a hash, not a wallet. One point is one cent billed. $BU0Y would count this sheet.
- [Full agent reference](https://bu0y.com/llms-full.txt): Complete API, auth, deposits, money rules, errors, Hermes YAML, model slugs
- [Discord](https://discord.gg/85ebxAjKgz): Support. Humans and agents, same channel.

## Product
- [Home](https://bu0y.com/): Harbor price current and get-started
- [Account](https://bu0y.com/account): Passkey login, balance, USDC deposit, API keys, usage
- [Models](https://bu0y.com/models): Full catalog; pass `model` yourself — clients do not auto-populate

## API (live)
- [GET /health](https://api.bu0y.com/health): Whether fills can happen right now (`ok`, `degraded`)
- [GET /v1/models](https://api.bu0y.com/v1/models): Union catalog, cheapest rate per model
- [POST /v1/quote](https://api.bu0y.com/v1/quote): Price a request across sources, with the range (free, no auth)
- [POST /v1/chat/completions](https://api.bu0y.com/v1/chat/completions): Prepaid fill. Bearer `bu0y_…` required
- [GET /v1/account](https://api.bu0y.com/v1/account): Remaining credits — `availableMicros` is what a new fill can draw. Bearer `bu0y_…`
- [GET /v1/harbor](https://api.bu0y.com/v1/harbor): OpenRouter retail vs Bu0y fill rates
- [GET /v1/stats](https://api.bu0y.com/v1/stats): Aggregate totals only
- [GET /v1/board](https://api.bu0y.com/v1/board): Anonymized spend ranks (handle + points). No account ids.

## Optional
- [OpenAI Chat Completions](https://platform.openai.com/docs/api-reference/chat): Client shape Bu0y mirrors
- [llms.txt specification](https://llmstxt.org/): Format this file follows