Open Chat

Free Forever AI Chat — a fast, private chatbot that anyone can use.

No sign-up, no API key, no model to choose. Open it and type. Conversations are written to your own browser, which is what makes an open URL safe to hand out: a visitor with no account leaves nothing behind to store, gate, or leak.

Sign in and the same chats follow you to your other devices. That is the only thing an account buys, and it is opt-in.

Open Chat is a single-deploy fullstack app — a React frontend and a handful of Netlify Functions, with no separate backend to host.

Browser (React + IndexedDB)          <-- source of truth
   |
   |-- /api/chat ------> Netlify streaming function --> gateway --> a model
   |-- /api/auth/* ----> Better Auth  ----------------> Postgres (optional)
   `-- /api/chats ------> chat history  --------------> Postgres (signed in only)

Why it is shaped this way

Chats are stored client-side first. A guest's transcripts never reach the deploy at all, which is what makes a public URL safe without building per-user quotas. Signing in adds a copy per chat so another device can pick it up; the browser stays the source of truth and the server is a mirror, not a backend.

Replies run on a streaming function. Netlify allows a streaming response 60 seconds on every plan; a plain synchronous function gets 10 on the free tier. Anything that talks to an LLM has to be the former.

Visitors configure nothing. There is no model picker and no place to paste a key. The server holds one credential and lets the gateway route each request, so there is no surface for a visitor to misconfigure and no support burden when they do.

Errors are written for visitors. "Set GATEWAY_API_KEY" is useless advice to someone who just wants an answer, and it leaks how the thing is wired. Operator detail goes to the function logs; the chat window gets a plain sentence and some idea of whether to retry.

Running it locally

Requires Node 20 or newer.

npm install
cp .env.example .env      # then fill in GATEWAY_API_KEY
npm run dev               # netlify dev on http://localhost:8888

netlify dev serves Vite and the functions on one origin, so /api/* behaves exactly as it does in production. npm run dev:vite runs the frontend alone, but the /api routes will 404.

Other scripts

Command	What it does
npm run build	Production build into dist/
npm run typecheck	Type-checks the frontend and the functions separately
npm run lint	ESLint, warnings treated as errors
npm run format	Prettier over the repo
npm run icons	Regenerates the PWA icons in public/
npm run auth:generate	Regenerates the Better Auth SQL schema
npm run auth:migrate	Applies the schema to a database you hold the URL to (safe to re-run)
npm run auth:schema	Regenerates auth-schema.sql from the installed better-auth

Deploying to Netlify

1. Push this repo and create a Netlify site from it. netlify.toml already sets the build command, publish directory, and functions directory.
2. Set the environment variables below under Site configuration → Environment variables.
3. Deploy. There is no second service to stand up.

Environment variables

Variable	Required	Notes
GATEWAY_BASE_URL	yes	OpenAI-compatible base URL, e.g. https://openrouter.ai/api/v1.
GATEWAY_API_KEY	yes	A key for that gateway. Put a spend limit on it — see below.
DEFAULT_MODEL	no	openrouter/auto routes per prompt; a concrete id pins one model.
FALLBACK_MODELS	no	Comma-separated chain tried on error or rate limit. Free models last.
MAX_TOKENS	no	Reply cap; keeps long answers inside the 60s streaming budget.
DATABASE_URL	no	Pooled Postgres URL. Enables accounts and cross-device chat history. Blank disables both; everything else keeps working.
BETTER_AUTH_SECRET	with auth	openssl rand -base64 32
BETTER_AUTH_URL	no	Netlify's URL is used when unset.
GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET	no	Enables the Google button.

One key pays for everyone. Chat has no sign-in gate, by design — that is the product. So every visitor spends the credit on GATEWAY_API_KEY. Set a hard spend limit on that key with your provider before the URL goes anywhere public; it turns the worst case into a capped number instead of an open tab. Ending FALLBACK_MODELS in free models helps too.

Turning on accounts

1. Create a free Postgres anywhere — Neon works well — and set its pooled connection string as DATABASE_URL in the Netlify site environment.
2. Set BETTER_AUTH_SECRET — openssl rand -base64 32.
3. Redeploy. /api/auth-status flips to {"enabled":true}.
4. For Google, add an OAuth client with redirect URI https://<your-site>/api/auth/callback/google, then set the two Google variables.

There is no migration step. Netlify treats DATABASE_URL as a secret, meaning the CLI can only ever read back a masked value — so nothing outside the deploy can reach the database to migrate it. The functions create their own tables on first use instead: netlify/lib/auth-schema.sql is idempotent, applied under an advisory lock, and skipped on later cold starts via a fingerprint row. npm run auth:migrate still exists for a local database you hold the URL to.

After upgrading better-auth, run npm run auth:schema to regenerate the auth tables from the installed version. It emits ADD COLUMN IF NOT EXISTS alongside each table, so a database created by an older release catches up rather than failing at runtime.

Using it

Action	How
Send	Enter (or Ctrl+Enter if you turn Enter-to-send off)
Newline	Shift+Enter
Attach an image	Click the image button, drop a file, or just paste a screenshot
Stop a reply	The stop button replaces send while streaming
Regenerate	Hover an assistant message → circular arrow
Edit and resend	Hover your own message → pencil. Everything after it is replaced.
Branch a chat	Hover any message → branch icon. Copies the thread up to that point.
Rename / pin / delete	The ... menu on any chat in the sidebar

Images are resized to fit within 1568px and re-encoded in the browser before they are sent, which keeps requests inside the function payload limit and costs fewer vision tokens.

Installing it as an app

Open Chat is a PWA. In Chrome or Edge, use the install icon in the address bar; on iOS, Share → Add to Home Screen. It opens in its own window and the shell is cached, so it starts instantly. When a new version deploys you get a toast offering to reload rather than being interrupted mid-sentence.

Project layout

netlify/
  functions/     chat.mts (streaming), health.mts, auth.mts, auth-status.mts
  lib/           settings, error mapping, SSE helpers, Better Auth setup + SQL
src/
  components/    UI, with shadcn primitives under components/ui
  hooks/         use-chat (the streaming engine), use-sessions, use-settings
  lib/           api, db (IndexedDB), sync, image (resizing), auth-client
  pages/         landing, about, auth, chat
scripts/         generate-icons.mjs — dependency-free PNG rasteriser for the PWA icons

Known limits

• A single reply must finish within 60 seconds. Very long outputs get a clear timeout message rather than a severed connection.
• Without an account, chats do not leave the browser they were typed in. That is the trade for not storing them.
• Attached images never sync. They are the bulk of a conversation by far, and this is a free deploy anyone can sign up to — so the metadata travels and the bytes stay put. A chat opened on another device shows that an image was there, not the image.
• Sync is last-write-wins per chat, not a merge. Editing the same chat on two devices between syncs loses the older side.
• Token usage is only shown when the upstream provider reports it; several do not.
• There is no model picker. Which model answered is reported per reply, but it is not yours to choose — that is deliberate, not missing.
• The UI components come from shadcn's radix-nova registry. Two files carry local modifications marked LOCAL MODIFICATION — ui/scroll-area.tsx (a viewport ref the streaming auto-scroll needs) and ui/sonner.tsx (no next-themes dependency). Re-apply them if you ever shadcn add those two.