One gateway. Every model. Uncensored.
A pool of 1,000 upstream keys behind one OpenAI-compatible endpoint.
Three jailbreak carriers, a scoring ledger, and a route that does not refuse.
What the gateway actually does
1,000 keys, rotated per request
Every call takes the next key from the pool and the next model in the class. The upstream cap is per model, not per key, so both rotate together.
Three jailbreak carriers, ranked
The ask is classified first, then the carriers for that class are ordered from the ledger. A refusal, or a prose answer to a code ask, is scored a loss — not a 200.
Every route is measured
Time to first token, escalation count and refusal count are written per request. The ranking is a number from that table, not a claim from a page.
From key to answer
Create a key
Register, open Tokens, and copy one token. That is the whole signup.
Swap the base URL
Leave the SDK alone. Replace the host with this gateway and send the same payload.
Read the completion
The carrier that answers is the one the ledger ranked first for that class. If it loses, the next one is tried before you ever see an error.
# the live counters, straight off the router curl https://lazarus-th.duckdns.org/health # {"ok":true,"models":["glm-5.3-flash","deepseek-v4-flash","minimax-m2.7"], # "median_ttft":36.95,"served":3,"escalated":0,"refused":0}
One key. The model you asked for.
Create a key, point your client at this gateway, and send the same payload.