The OpenAI-compatible API that runs 18+ models

An OpenAI-compatible API is an HTTP service that mimics the request and response shapes of the OpenAI API. You point an OpenAI SDK at its base_url, and your code runs unchanged. Plugsky is one of these — but it also runs 18+ models behind the same endpoint, not just one vendor's weights.

This page explains exactly what Plugsky's OpenAI-compatible surface covers, how to migrate, and the gotchas to avoid.

What "OpenAI-compatible" actually means

OpenAI's API is a de-facto standard. The /v1/chat/completions endpoint takes a JSON body with a model, messages, and a handful of parameters, and returns a JSON response with a choices array. Hundreds of SDKs and tools — Cursor, Continue, Aider, LangChain, LlamaIndex, Vercel AI SDK, the official OpenAI libraries — speak this shape.

Plugsky replicates the surface 1:1. The endpoint URLs, request bodies, response shape, error codes, and streaming format all match. The only thing that changes is the model names and the price.

Endpoints Plugsky implements

Supported SDKs and languages

Anything that can talk HTTPS works with Plugsky. Tested daily:

• Python: openai ≥ 1.0, httpx, requests
• Node.js / TypeScript: openai, vercel/ai-sdk, langchainjs, llamaindex
• Browser: fetch, openai-fetch
• Go: sashabaranov/go-openai
• Rust: async-openai
• Ruby: ruby-openai
• PHP: openai-php/laravel
• CLI: curl, httpie

One-line migration

# Before (OpenAI)
client = OpenAI(api_key="sk-...")

# After (Plugsky — only base_url changes)
client = OpenAI(
    api_key="sk-live-PLUGSKY...",
    base_url="https://api.plugsky.com/v1",
)

resp = client.chat.completions.create(
    model="plugsky-pro",
    messages=[{"role":"user","content":"Hello!"}],
)

The same applies to every other SDK. Set the base URL to https://api.plugsky.com/v1 and you're done.

OpenAI parameters Plugsky honors

• temperature, top_p, frequency_penalty, presence_penalty
• max_tokens, stop, seed
• response_format={"type":"json_object"} for JSON mode
• tools, tool_choice for function calling
• stream=true for SSE streaming
• user for per-end-user tracking
• logprobs and top_logprobs on supported models

Limitations vs. native OpenAI

You do not get: DALL·E, the Assistants API shape, the Threads / Messages / Runs endpoints, fine-tuning, or OpenAI's built-in moderation endpoint. For those, run OpenAI alongside Plugsky. Most teams use Plugsky for chat / completion / embeddings and keep OpenAI for image gen and TTS.

Request limits & error shapes

Every response from the API — success or failure — is JSON with the OpenAI error schema: {"error":{"message","type","code","param"}}. You will never receive an HTML page from an API endpoint; HTML pages only exist on the marketing/docs site.

Request size: the maximum request body is 16 MB (this covers a full 128K-token context plus tool definitions and file excerpts). Requests above 16 MB return 413 request_too_large.

Context windows: all chat models advertise a 128K context window; the upstream service enforces a 1,000,000-token ceiling per request. If your agent compacts long conversations, keep the total request under these limits — 400 errors include the exact token counts so you can size your compaction strategy.

Streaming: errors during streaming are returned as JSON before the SSE stream starts (non-200 status). Once streaming begins you receive OpenAI-format SSE chunks ending in data: [DONE].

Frequently asked questions