ChatGPT Subscription

If you have a ChatGPT Plus or Pro subscription, you can run your assistant-ui app against it locally instead of paying per token with an API key. OpenAI's Codex CLI authenticates via OAuth and caches tokens on your machine; requests made with those tokens hit the Responses API endpoint at chatgpt.com and bill to your plan. The same mechanism powers Codex itself and a growing set of local agents.

This is a personal-use setup for apps running on your own machine. For anything deployed or shared, use an API key.

Log in with Codex

Authenticate once with the Codex CLI. The login flow runs OAuth in your browser and writes access and refresh tokens to ~/.codex/auth.json:

npx @openai/codex login

Everything below reads that file; no environment variables or API keys are involved.

AI SDK route

The community AI SDK provider openai-oauth-provider picks up ~/.codex/auth.json automatically, refreshes tokens as needed, and supports streaming, tool calls, and reasoning traces.

npm install openai-oauth-provider

In your chat route, swap @ai-sdk/openai for the OAuth provider. Starting from the default assistant-ui template route, only the model changes; frontendTools and the UI message stream response work as before:

import { createOpenAIOAuth } from "openai-oauth-provider";
import { frontendTools } from "@assistant-ui/react-ai-sdk";
import {
  type JSONSchema7,
  streamText,
  convertToModelMessages,
  type UIMessage,
} from "ai";

const openai = createOpenAIOAuth();

export async function POST(req: Request) {
  const {
    messages,
    system,
    tools,
  }: {
    messages: UIMessage[];
    system?: string;
    tools?: Record<string, { description?: string; parameters: JSONSchema7 }>;
  } = await req.json();

  const result = streamText({
    model: openai("gpt-5.3-codex"),
    messages: await convertToModelMessages(messages),
    system,
    tools: {
      ...frontendTools(tools ?? {}),
    },
  });

  return result.toUIMessageStreamResponse({
    sendReasoning: true,
  });
}

The provider discovers which models your plan includes (for example gpt-5.4 and gpt-5.3-codex) from your ChatGPT account.

Eve agents

Eve's defineAgent accepts a provider-authored AI SDK LanguageModel, so the same provider drops into an eve agent config:

import { defineAgent } from "eve";
import { createOpenAIOAuth } from "openai-oauth-provider";

export default defineAgent({
  model: createOpenAIOAuth()("gpt-5.3-codex"),
});

OpenAI-compatible proxy

For stacks that take an OpenAI base URL rather than an AI SDK model, the same project publishes a sibling package, openai-oauth, whose CLI runs a local proxy:

npx openai-oauth
# OpenAI-compatible endpoint ready at http://127.0.0.1:10531/v1

Point any OpenAI-compatible client at that URL. No real API key is involved, but most clients require one at construction, so pass a placeholder:

import { OpenAI } from "openai";

const openai = new OpenAI({
  baseURL: "http://127.0.0.1:10531/v1",
  apiKey: "unused",
});

The proxy binds to 127.0.0.1 by default; keep it that way, since any caller that can reach the endpoint bills your subscription.