Four lines to Qwen.

Overview

Kalla is an inference gateway. You talk to one OpenAI-compatible surface; we handle the upstream credentials, quotas, streaming normalisation and the usage ledger. Eight models are live today:

Quickstart

1. 1Get an API key

   Keys are issued by the Kalla team and shown once in the dashboard. They look like kalla_live_xxxxxxxxxx_… and are stored hashed — if you lose one, rotate it instead of asking for a copy.

2. 2Point your client at the gateway

   Replace the base URL and keep the OpenAI SDK you already use.

   app.ts

   Copy

   import OpenAI from 'openai';
   
   const kalla = new OpenAI({
     apiKey: process.env.KALLA_API_KEY,
     baseURL: 'https://api.kalla.website/v1',
   });

3. 3Send your first completion

   first-call.ts

   Copy

   const completion = await kalla.chat.completions.create({
     model: 'Qwen3.8',
     messages: [
       { role: 'system', content: 'Answer in the user language.' },
       { role: 'user', content: 'اشرح الفرق بين SSE و WebSocket بجملتين' },
     ],
   });
   
   console.log(completion.choices[0]?.message?.content);

Authentication

Send your key as a bearer token. x-api-key is accepted for Anthropic-style clients. Keys are scoped (chat, models), can expire, and are checked on every request — revocation takes effect immediately, with no token to wait out.

export KALLA_API_KEY="kalla_live_xxxxxxxxxx_your_secret"

# preferred
curl https://api.kalla.website/v1/models -H "Authorization: Bearer $KALLA_API_KEY"

# also accepted
curl https://api.kalla.website/v1/models -H "x-api-key: $KALLA_API_KEY"

Endpoints

Request

{
  "model": "Qwen3.8",
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "Summarise our Q3 report." }
  ],
  "temperature": 0.7,
  "max_tokens": 1024,
  "stream": false
}

Response

{
  "id": "chatcmpl-…",
  "object": "chat.completion",
  "created": 1770000000,
  "model": "Qwen3.8",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "…" },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 180,
    "total_tokens": 204
  },
  "kalla": {
    "model": "Qwen3.8",
    "provider": "primary",
    "latency_ms": 812,
    "request_id": "…"
  }
}

kalla is additive metadata; OpenAI clients ignore unknown fields. Reference request_id in support tickets — it appears in every error too.

Streaming

Set stream: true and read SSE frames. The stream always ends with data: [DONE]. Add stream_options.include_usage to receive a final chunk with token counts.

data: {"id":"chatcmpl-…","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

data: {"id":"chatcmpl-…","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"بغداد"},"finish_reason":null}]}

data: {"id":"chatcmpl-…","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]

Disable proxy buffering when self-hosting (nginx: proxy_buffering off;). We already send X-Accel-Buffering: no. If the upstream fails mid-stream, an error object is written as a final SSE frame.

Errors

Every error is OpenAI-shaped and carries a stable machine code plus a request id.

{
  "error": {
    "message": "Rate limit reached: 60 requests per minute for this key.",
    "type": "rate_limit_error",
    "code": "rate_limit_exceeded",
    "request_id": "9f2c…"
  }
}

Rate limits & quotas

Responses include RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset (plus X- aliases). On 429, wait for Retry-After seconds. Use exponential backoff with jitter for 502/503.

• Every plan includes the full catalogue, streaming and tool calling. There is no per-token metering.
• Fair-use rate limits apply per API key, so a runaway script cannot exhaust the fleet for everyone else.
• When the shared GPUs are busy, requests wait in a short queue and responses can be slower — nothing is dropped.
• Need more throughput or dedicated capacity? Talk to us and we will size a plan for your team.