unnma.Docs
Get started

Welcome to Unnma

Unnma is a single API that routes your inference to nine AI vendors and applies a proprietary optimization layer called COP — Cognitive Orientation Prompting. You write code against the Anthropic or OpenAI SDK you already use; you change one line (the base URL); your calls land at api.unnma.ai; we forward them to the underlying provider with COP applied. Coding and Customer Support workloads see roughly 24% fewer output tokens and a 60% reduction in latency on average. Across mixed categories, the average is closer to 12%. Your bill shrinks accordingly.

How it works

python
from anthropic import Anthropic
 
client = Anthropic(
    api_key="unnma-sk-...",
    base_url="https://api.unnma.ai/v1",
)
 
response = client.messages.create(
    model="anthropic/claude-opus-4-7",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, world."}],
)

Three things are happening:

  1. You changed your base URL. Everything else in your code is unchanged — same SDK, same request shape, same response shape.
  2. You used a prefix-qualified model name. anthropic/claude-opus-4-7 tells us which vendor to route to. The OpenAI SDK works the same way against /v1/chat/completions.
  3. COP is applied to your request before forwarding. The result is a measurably lower token count and latency.

You get the savings as a lower charge on the request. Your dashboard shows the per-call delta — what you paid through us versus an estimate of what the same call would have cost going direct.

Why this exists

Two problems sit next to each other in production AI work:

  • LLM inference is expensive Our innovative inference optimization technique, COP, reduces costs by up to 70% per request.
  • **LLM inference can be slow ** COP reduces latency by 60% on average for coding and customer support tasks.
  • Vendor lock-in is real. Switching from Claude to GPT to Gemini means rewriting your client code and re-testing every prompt. Most teams don't switch even when the economics would say to.

COP Gateway addresses both at once. The vendor routing happens at the model-name prefix — anthropic/..., openai/..., groq/..., gemini/..., plus four more direct integrations and OpenRouter as a long-tail fallback. COP runs automatically on every request. You don't change SDK; you don't change your prompts; you just save money on every call.

What's in these docs

  • Quickstart — get a key, make your first call. Five minutes.
  • Authentication — how API keys work, how to create and revoke them.
  • Pricing — the 9% markup on inference, the service fee on top-ups, the welcome credit.
  • /v1/messages reference — the Anthropic-shape endpoint. Use this if you were using the Anthropic SDK.
  • /v1/chat/completions reference — the OpenAI-shape endpoint. Use this if you were using the OpenAI SDK.
  • Error codes — every status code, every error type, what to do when you see it.
  • Rate limits — per-key and per-account limits, daily spend caps, the Retry-After pattern.

Next

Start at the Quickstart. It is the fastest path to a working call.