onexai — Python SDK for the 1xAi gateway

1xAi is an OpenAI-compatible AI gateway for Iran: one key and one base URL for OpenAI, Anthropic Claude, Google Gemini and DeepSeek — no VPN, no foreign card, billed in Toman.

pip install 1xai-sdk

from onexai import OneXAI, cost_of_response

client = OneXAI()  # reads ONEXAI_API_KEY, falls back to OPENAI_API_KEY

reply = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "سلام! یک جوک فارسی بگو"}],
)
print(reply.choices[0].message.content)
print(cost_of_response(reply))     # -> 102 Toman (gpt-4o-mini, 14in/23out)

Why this package is thin (and why that is the point)

1xAi's whole claim is that it is the OpenAI API at a different base URL. So this package does not re-implement an SDK — it depends on the official openai package and pre-configures it. Writing our own HTTP, streaming, retry and tool-call layer would mean shipping a worse copy of a client OpenAI already maintains, and it would drift every time the API gains a field.

You do not strictly need this package at all:

from openai import OpenAI
client = OpenAI(api_key="1xai-...", base_url="https://1xai.ir/v1")

That is the entire integration. What onexai adds is the small set of things the generic SDK cannot know about a gateway in Iran:

CLI

onexai doctor                        # diagnose key, connectivity, credit, upstream health
onexai price gpt-4o-mini 1000 500    # Toman cost of a call
onexai models --provider anthropic   # what you can call, with prices
onexai status                        # live upstream health, as seen from Iran

Cost estimation

Prices come from https://1xai.ir/api/models — public, no key required. A snapshot ships with the package so the helper works offline; pass live=True to fetch current prices.

from onexai import estimate_cost

e = estimate_cost("claude-opus-4-5", input_tokens=50_000, output_tokens=2_000)
print(e.total_toman)   # 66,067.0
print(e.provider)      # anthropic

Audio and image endpoints are billed per second and per image, so they are not in the token catalogue. estimate_cost("whisper-1", ...) raises UnknownModelError explaining that, rather than quietly returning zero.

Native passthroughs

The OpenAI-compatible path silently drops provider-specific features. When you need them, use the native endpoints:

from onexai import anthropic_client       # pip install '1xai-sdk[anthropic]

claude = anthropic_client()               # -> https://1xai.ir/anthropic
msg = claude.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    system=[{"type": "text", "text": BIG_DOC,
             "cache_control": {"type": "ephemeral"}}],   # dropped on /v1
    messages=[{"role": "user", "content": "خلاصه کن"}],
)
print(msg.usage.cache_read_input_tokens)

Full cookbook

Runnable recipes for Cursor, Continue, Cline, Open WebUI, n8n, streaming, Persian Whisper transcription, image generation and cost control live in the repository README.

License

MIT. Not an official 1xAi product unless the 1xAi team adopts it; see the repository.