An OpenAI-compatible gateway in front of many inference providers. One key, one balance, per-request budgets that actually stop runaway agents — and a raw stream that never buffers.
# one key, ordinary OpenAI SDK from openai import OpenAI client = OpenAI( base_url="https://ai.alcov.space/v1", api_key="sk-ai-...", ) stream = client.chat.completions.create( model="venice/deepseek-v4-1-flash-1017-uncensored", messages=[{"role": "user", "content": "hello"}], stream=True, )
Every request reserves credit before the model runs, clamps max_tokens to what the balance covers, then settles on real usage. A looping agent gets a 402, not an invoice.
Cheapest, fastest, or throughput-per-dollar. Candidates are tried in order with the hold released on each failure.
Responses are piped straight through. Usage is read from the final SSE chunk so billing is exact, not estimated.
No markup on censored models. You pay exactly what Vercel AI Gateway charges — zero platform fee on tokens.
Top up once, spend it across every provider. Balance never expires, and keys, caps and the playground draw on the same pot.
Uncensored and abliterated builds carry the provider's own price. No filtering layer sits between you and the reply.
Frontier models, open weights, and refusal-free builds behind a single credential. Adding a provider is a catalog entry, not a migration.