OpenAI-compatible recipes — connect any model in ≤10 lines

Status: A5 (S, 2026-06). The recommended path for any new model integration. Existing model-* packages keep working but are now documented as presets rather than the primary integration story.

The Mastra leaderboard race (94 providers / 3300+ models, March 2026) and the AI SDK 6 unified gateway() API both proved one thing: hand-rolling provider adapters is a losing race. Almost every model worth using — OpenAI, Anthropic-via-OpenRouter, Mistral, DeepSeek, Qwen, Doubao, Moonshot, Zhipu, MiniMax, Together, Fireworks, Groq, Anyscale, Ollama, LM Studio, vLLM, llama-server — speaks an OpenAI-compatible /chat/completions.

wasmagent's answer: GenericOpenAICompatModel in @wasmagent/core. One concrete class, three constructor args, every provider's quirks expressed as runtime config. New providers become README recipes (this file), not new packages.

import { GenericOpenAICompatModel } from "@wasmagent/core";

Recipe — Ollama / LM Studio (local, no API key)

const model = new GenericOpenAICompatModel("qwen2.5:14b", "http://localhost:11434/v1", {
  apiKey: "ollama", // any non-empty string; Ollama ignores it
  extraCapabilities: { localEndpoint: true, metered: false },
});

For LM Studio swap base URL to http://localhost:1234/v1. For llama-server / vLLM: http://localhost:<port>/v1.

Recipe — OpenRouter (every model on one URL)

const model = new GenericOpenAICompatModel(
  "anthropic/claude-3.5-sonnet",
  "https://openrouter.ai/api/v1",
  {
    apiKey: process.env.OPENROUTER_API_KEY,
    defaultHeaders: {
      "HTTP-Referer": "https://your-app.example",
      "X-Title": "your-app",
    },
  }
);

OpenRouter's full catalogue (~300 models) is available the same way — just change modelId.

Recipe — Vercel AI Gateway

const model = new GenericOpenAICompatModel("openai/gpt-4o-mini", "https://gateway.ai.vercel.app/v1", {
  apiKey: process.env.VERCEL_AI_GATEWAY_KEY,
});

Recipe — Cloudflare AI Gateway

const model = new GenericOpenAICompatModel(
  "openai/gpt-4o-mini",
  `https://gateway.ai.cloudflare.com/v1/${ACCOUNT_ID}/${GATEWAY_NAME}/openai/v1`,
  { apiKey: process.env.OPENAI_API_KEY }
);

Recipe — DeepSeek (with reasoning_content round-trip)

DeepSeek-R1 emits reasoning_content on a non-standard delta field and expects it echoed back on tool-result turns. One option flag covers both:

const model = new GenericOpenAICompatModel("deepseek-r1", "https://api.deepseek.com/v1", {
  apiKey: process.env.DEEPSEEK_API_KEY,
  reasoningContentField: "reasoning_content",
  reasoningRoundTrip: "tool-turns-only",
});

This is exactly what @wasmagent/model-deepseek does internally — it's preserved as a named preset, but the recipe above gets you 100% of the behaviour without an extra npm install.

Recipe — Groq (super-fast Llama / Mixtral)

const model = new GenericOpenAICompatModel("llama-3.3-70b-versatile", "https://api.groq.com/openai/v1", {
  apiKey: process.env.GROQ_API_KEY,
});

Recipe — Together / Fireworks

// Together
new GenericOpenAICompatModel("meta-llama/Llama-3.3-70B-Instruct-Turbo", "https://api.together.xyz/v1", {
  apiKey: process.env.TOGETHER_API_KEY,
});

// Fireworks
new GenericOpenAICompatModel(
  "accounts/fireworks/models/llama-v3p3-70b-instruct",
  "https://api.fireworks.ai/inference/v1",
  { apiKey: process.env.FIREWORKS_API_KEY }
);

Recipe — extraRequestParams & extraThinkingParams for one-off quirks

Some endpoints accept non-standard fields like enable_thinking (Qwen3) or effort levels exposed under different names. Pass them through without subclassing:

const model = new GenericOpenAICompatModel("qwen3-235b-instruct", "https://dashscope.aliyuncs.com/compatible-mode/v1", {
  apiKey: process.env.DASHSCOPE_API_KEY,
  extraRequestParams: { enable_thinking: true },
  reasoningContentField: "reasoning_content",
  reasoningRoundTrip: "tool-turns-only",
});

When you DO need a model-* package

If your provider differs in ways GenericOpenAICompatModel cannot express in options (e.g. a non-streamed reasoning protocol, custom request shape, non-OpenAI streaming envelope), subclass OpenAICompatModel directly and override the relevant protected hooks. That is what the existing model-deepseek, model-doubao, model-moonshot, model-zhipu, model-qwen, model-minimax packages do. We are not removing them — they remain as imported-by-name presets and as canonical examples for contributors. We are just no longer growing this list as the primary integration story.

See also

• packages/core/src/models/OpenAICompatModel.ts — the implementation. The GenericOpenAICompatModel class is at the bottom of the file.
• docs/guides/code-mode.md — uses these models inside the code-mode MCP server pattern.