noumen — coding agents are products. This is the library.

Setup

pnpm add @ai-sdk/openai

import { AiSdkProvider } from "noumen";
import { createOpenAI } from "@ai-sdk/openai";

const openai = createOpenAI({
  apiKey: process.env.OPENAI_API_KEY!,
  baseURL: "https://...",          // optional, for Azure / proxies / compatible APIs
});

const provider = new AiSdkProvider({
  // `.chat(id)` pins the call to the chat/completions endpoint.
  // Drop `.chat(...)` to use the Responses API instead.
  model: openai.chat("gpt-5"),
});

Options

All connection-level options (API key, base URL, custom headers, organization, project) come straight from createOpenAI. noumen does not re-export them — whatever the AI SDK ships, you can pass.

Option	Source	Description
`apiKey`	`createOpenAI`	OpenAI API key.
`baseURL`	`createOpenAI`	Override the API base URL (for Azure, proxies, or compatible APIs).
`headers`	`createOpenAI`	Custom headers sent with every request (useful for proxies, auth, or tracking).
`organization`, `project`	`createOpenAI`	OpenAI organization and project IDs.

Compatible APIs

The baseURL option lets you use any OpenAI-compatible API. This includes:

Azure OpenAI — point to your Azure endpoint.
OpenRouter — or use the dedicated OpenRouter page for rank-tracking headers.
Local models — Ollama, vLLM, or other local servers.

import { AiSdkProvider } from "noumen";
import { createOpenAI } from "@ai-sdk/openai";

const openrouter = createOpenAI({
  apiKey: process.env.OPENROUTER_API_KEY!,
  baseURL: "https://openrouter.ai/api/v1",
});

const provider = new AiSdkProvider({
  model: openrouter.chat("anthropic/claude-sonnet-4"),
});

createOpenAI()(modelId) returns a Responses API model by default. That endpoint is the most feature-rich (structured output, stateful tool use, image generation), but some OpenAI-compatible proxies only implement chat/completions. When you're pointing at a proxy, pin to chat/completions:

const provider = new AiSdkProvider({ model: openai.chat("gpt-5") });

Use the default openai("gpt-5") for the Responses API.

Reasoning / thinking

GPT-5 and o-series models support reasoning effort. Pass it on AgentOptions.thinking (mapped to reasoning_effort: "high") or directly via reasoningEffort: "low" | "medium" | "high" | "minimal" on ChatParams:

const agent = new Agent({
  provider,
  sandbox: LocalSandbox({ cwd: "." }),
  options: { thinking: { type: "enabled", budgetTokens: 10_000 } },
});

Streaming

AiSdkProvider uses stream: true and surfaces include_usage metadata automatically. No additional configuration is needed.

Models

Any model available through the OpenAI API works. Common choices:

gpt-5 — latest flagship
gpt-4o — fast, capable, good default
gpt-4.1 — previous generation
gpt-4o-mini — cheaper, faster for simpler tasks
o3-mini / o4-mini — reasoning models

OpenAI