the open sea — routed routing live
every token is cargo. we run the manifests. fig. 01 — the crossing
openai-compatible · one endpoint · payment channels

The inference layer that pays for itself.

One endpoint in front of every model — routed to the cheapest capable provider, failed over when one blinks, metered per call against a funded payment channel. No subscription, no markup — just cheaper, more reliable inference.

↓
01 — route

The cheapest capable model. Every call.

Auto-tiering per request, circuit breakers per provider, a failover receipt on every response. Typically 70%+ under baseline — and your own model stays yours; routing is opt-in.

four stepsfrom wallet to answer
1 · connecta wallet you already havePhantom · Solflare · PayBox
2 · talkrouted to the cheapest capable modelauto
3 · drawevery call meters off-chain, to the tokenper call
4 · your boxlocal models run free first$0.00
02 — reliability

When a provider blinks, you don't.

Circuit breakers per provider, health tracking, and automatic failover to the next rung — with the full receipt on every response. Your agent never sees a provider outage; it just sees an answer.

Request model: auto and the gateway tiers, costs, and fails over for you — pinned models stay pinned, and local Ollama traffic always runs free.

failover receiptsx_shipyard
openai → nouscircuit open312msok
gpt-5-mini → nanotier: refactor−71%ok
local / ollamafree tier$0.00free
03 — models

One endpoint. Every model.

Drop-in for anything that speaks OpenAI — Claude Code, your own agents, plain curl. Anthropic, OpenAI, OpenRouter, UsePod, and Nous/Hermes behind one interface, plus your local models for free. Keep your own model and API keys; routing is strictly opt-in.

# issues an sk-shipyard key + a status line — model untouched npx shipyard-inference connect --url https://shipyard-inference.vercel.app
from openai import OpenAI client = OpenAI( base_url="https://shipyard-inference.vercel.app/v1", api_key="sk-shipyard-…", # your key from `connect` ) resp = client.chat.completions.create( model="auto", # cheapest capable model, chosen per call messages=[{"role": "user", "content": "hello from my agent"}], ) print(resp.choices[0].message.content)
# or curl it directly curl https://shipyard-inference.vercel.app/v1/chat/completions \ -H "authorization: Bearer ***" \ -H "content-type: application/json" \ -d '{"model":"auto","messages":[{"role":"user","content":"hi"}]}'
04 — pricing

Fund a channel. Settle once.

No subscription, no seat fees, no markup. Fund a payment channel once in USDC — every call then draws it down off-chain at the provider's own rate, and closing the channel settles the whole session in one on-chain transaction. Spend ceilings per key so a runaway agent can't drain you.

runs where you wantno middleman
your box / ollamalocal models$0.00free
cloud burstwhen local isn't enoughautorouted
open channelfund once, USDC escrowedx402upto
close & settleone on-chain settle for the draw1 txsettle
05 — connect

Change one URL. Keep your model.

# issues an sk-shipyard key + a status line — model untouched npx shipyard-inference connect --url https://shipyard-inference.vercel.app