<!-- Source: https://docs.khwan.ai/chatgpt -->
<!-- The whole documentation in one file: https://docs.khwan.ai/llms-full.txt -->

# ChatGPT (Custom GPT)

Use Khwan as durable memory inside ChatGPT — on the web, **free tier included**.
Khwan never runs a model; in ChatGPT, **ChatGPT is the model**. Khwan is the memory
it forgets between chats. Each person signs in with their own account, so memories
are isolated per user.

  Free ChatGPT accounts can **use** a published Custom GPT (they just can't build
  one). MCP connectors need Plus/Pro — a Custom GPT + OAuth Action does not, which
  is why this path reaches free users.

## For users

1. Open the **Khwan** GPT on [chatgpt.com](https://chatgpt.com) (works in any browser).
2. The first time it needs memory, a **“Sign in to Khwan”** button appears — click it,
   authorize once, and you're back in the chat.
3. That's it. Ask things; Khwan quietly recalls what matters and remembers new facts.
   The same brain follows you into the [dashboard](https://app.khwan.ai) and any
   other host you connect (it's linked by your verified email).

## The loop

Under the hood the GPT runs Khwan's pure-memory loop — the same one as every other
host (see [Bring Your Own Model](/byom)):

- **`prepareTurn(input)`** → memory-grounded context for your message. No LLM call.
- ChatGPT answers, grounded in that context.
- **`recordAnswer(turn_token, answer)`** → Khwan persists + learns; the next prepare
  is sharper. Always pass the exact `turn_token` back — never invent one.

---

## For operators — publishing the GPT

Setting this up yourself. The only moving parts are a managed IdP (for per-user
OAuth) and the GPT itself; the engine change is minimal.

### Configure the IdP (WorkOS AuthKit)

Khwan uses [WorkOS AuthKit](https://workos.com/) as the OAuth 2.1 authorization
server (free to ~1M MAU; Auth0 works as a fallback).

- Create an **OAuth Application (third-party)** → note the **Client ID + Secret**.
- Enable **Google** + email/passwordless login, and the `offline_access` scope (refresh).
- In the **JWT template**, include `email` and `email_verified` — Khwan links a user's
  ChatGPT and dashboard logins by *verified* email, so both reach one brain.
- Note the **Issuer**, **JWKS URL**, and (optionally) **Audience**.

### Point the engine at the IdP

Set on the engine (absent → the OAuth path is simply disabled):

```bash
KHWAN_OIDC_ISSUER=https://your-tenant.workos.com
KHWAN_OIDC_JWKS_URL=https://your-tenant.workos.com/oauth2/jwks
KHWAN_OIDC_AUDIENCE=khwan-api        # optional; verified only when set
```

The engine then accepts the IdP's access token as `Authorization: Bearer <jwt>`,
verifies it offline against the JWKS, and auto-provisions a free account per user.

### Build the Custom GPT

- In the GPT editor, add an **Action** and paste
  [`chatgpt-action.yaml`](https://github.com/khwanlabs/khwan-openapi) (prepare /
  record / memory, trimmed for ChatGPT).
- Set **Authentication → OAuth**: paste the Client ID/Secret, the Authorization URL
  (`…/oauth2/authorize`), Token URL (`…/oauth2/token`), and scope `khwan.memory`.
- Give it these instructions:

```text
You have durable memory via Khwan. At the start of a task, or whenever you need a
fact that may have come up before, call prepareTurn(input) and ground your answer in
the returned context. After you answer, when a durable fact, preference, or decision
emerges, call recordAnswer(turn_token, answer) using the EXACT turn_token from the
matching prepareTurn — never invent one. Do not put secrets, passwords, or sensitive
personal data into memory.
```

### Publish

Verify the `khwan.ai` domain in your OpenAI builder profile, then publish. GPTs with
OAuth actions are usually **reviewed by OpenAI**, so allow a little lead time.

  **Privacy.** A public free GPT invites strangers. Keep prod rate limiting on
  (`KHWAN_REDIS_URL`), and cap the synthesis cost of auto-provisioned free accounts
  with `KHWAN_FREE_SYNTHESIS_DISABLED=1` (or leave free on its weekly cadence). Tell
  users, in the consent screen, not to store sensitive personal data.