Integrations

Anything that accepts a custom OpenAI base URL works. These are the exact settings for the tools people ask about most.

In every case the two values are the same: base URL https://api.infersia.com/v1 and your isk-v1-… key.

Cline (VS Code)

Open the Cline settings, choose OpenAI Compatible as the API provider, and fill in:

Base URL:  https://api.infersia.com/v1
API Key:   isk-v1-...
Model ID:  qwen/qwen3-8b

Continue

~/.continue/config.jsonjson

{
  "models": [
    {
      "title": "Qwen3 8B (Infersia)",
      "provider": "openai",
      "model": "qwen/qwen3-8b",
      "apiBase": "https://api.infersia.com/v1",
      "apiKey": "isk-v1-..."
    }
  ]
}

Aider

export OPENAI_API_BASE=https://api.infersia.com/v1
export OPENAI_API_KEY=isk-v1-...

aider --model openai/qwen/qwen3-8b

OpenWebUI

Under Settings → Connections → OpenAI API, add a connection with the base URL and your key. Our /v1/models endpoint populates the model list automatically.

LangChain

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="qwen/qwen3-8b",
    base_url="https://api.infersia.com/v1",
    api_key="isk-v1-...",
    streaming=True,
)

LlamaIndex

from llama_index.llms.openai_like import OpenAILike

llm = OpenAILike(
    model="qwen/qwen3-8b",
    api_base="https://api.infersia.com/v1",
    api_key="isk-v1-...",
    is_chat_model=True,
    context_window=32768,
)

LiteLLM

config.yamlyaml

model_list:
  - model_name: qwen3-8b
    litellm_params:
      model: openai/qwen/qwen3-8b
      api_base: https://api.infersia.com/v1
      api_key: os.environ/INFERSIA_API_KEY

Vercel AI SDK

import { createOpenAI } from '@ai-sdk/openai';
import { streamText } from 'ai';

const infersia = createOpenAI({
  baseURL: 'https://api.infersia.com/v1',
  apiKey: process.env.INFERSIA_API_KEY,
});

const result = streamText({
  model: infersia('qwen/qwen3-8b'),
  prompt: 'Explain KV caching.',
});

Agent loops and prefix caching

If your framework replays a long system prompt on every iteration — which most agent loops do — keep that prefix byte-identical between calls. Our endpoints have prefix caching enabled and the discount is passed through, so a stable prefix bills at a quarter of the normal input rate.

Anything that changes the leading bytes defeats it: a timestamp in the system prompt, tools listed in a different order, a session id inserted at the top. Move variable content to the end of the prompt and the cache does the rest. Your activity log shows the cached token count per request, so you can verify it’s working rather than assuming.