NextChat

Last reviewed Aug 26, 2026 · 810 words

NextChat is a client, not a platform — understand that and every design choice makes sense. Formerly ChatGPT-Next-Web, it is an 88,000-star MIT-licensed chat interface that talks to model providers you bring keys for: OpenAI, Anthropic, Gemini, and most OpenAI-compatible endpoints. There is no user database, no admin panel, no document pipeline — which is exactly why it deploys in minutes, idles in 512 MB of RAM, and earns its "Easy" difficulty rating in the catalogue. If you want a household AI portal with accounts and RAG, look elsewhere. If you want the fastest possible ChatGPT-style UI over keys you control, this is it.

From zero to chatting in one command

The Docker path:

docker run -d -p 3000:3000 \
  -e OPENAI_API_KEY=sk-your-key \
  -e CODE=pick-a-long-access-password \
  yidadaa/chatgpt-next-web

Open port 3000 and you are talking to GPT with no further setup. The even shorter path is the one-click Vercel deploy from the NextChat repository — fork, set two environment variables, done, hosted free. Purists will note Vercel is not self-hosting, and they are right; I treat it as the try-before-you-commit route and run the container at home for the real thing. Either way, first chat happens inside 10 minutes, which no other tool in the AI chat UI category reliably matches.

CODE is the only lock on the door

That CODE variable is an access password, and skipping it while exposing the app publicly means anyone who finds the URL chats on your API key — at your expense, against your provider's terms, until you notice the bill. This is the single most common NextChat mistake, because the app runs fine without it. Set a long value from day one, and treat it for what it is: one shared password, not authentication. There are no per-user accounts, no rate limits, no audit trail; the catalogue tags it single-user for a reason. I share my instance with exactly one other person I would trust with the key itself. For anything beyond that circle, put real auth (a reverse proxy with SSO) in front, or pick a multi-user tool instead.

More client surface than you expect

For a "lightweight" app, the feature list is generous: prompt templates via a built-in mask library, per-conversation model and temperature settings, automatic context compression to stretch long chats, markdown with code highlighting, and export of conversations. Because it began life as a bring-your-own-key ChatGPT front end, it also fills the gap that pushed many people off the official apps: one interface, several providers, no $20/month subscription — you pay per token instead, which for light use runs a few dollars a month. The cross-platform story is real too, with desktop builds for Windows, macOS, and Linux alongside web and mobile, all one codebase talking to the same providers.

Local models work, with one URL

NextChat speaks the OpenAI API dialect, so anything that serves that dialect can stand in for OpenAI. Point the custom endpoint at an Ollama instance (or LocalAI, or any OpenAI-compatible server) and the same UI drives fully local models — worth doing for private notes-adjacent chats even if you keep cloud models for hard problems. Two caveats from running this setup: conversations live in browser storage unless you configure sync, so they are per-device by default, and NextChat does no model management — Ollama pulls and serves the models, NextChat is only ever the front end. That division of labor is fine; just know which side owns which problem.

Where Open WebUI takes over

The moment you need user accounts, document upload with retrieval, or admin control over who uses which model, you have outgrown NextChat's design. Open WebUI carries all of that at the cost of a heavier footprint and more setup, and the full comparison is in NextChat vs Open WebUI. The two are not really rivals so much as adjacent points on a weight/capability curve: NextChat for individuals with API keys, Open WebUI for households and teams around shared local models.

What I'd do

Run the Docker container behind your existing reverse proxy, CODE set to 20+ random characters, keys for one cloud provider plus a custom endpoint aimed at Ollama. Use it as your daily driver for a month; total cost will be single-digit dollars in tokens. If nobody else ever needs an account and you never miss document chat, you have found your permanent setup — I know several people who did. If you do start missing those things, the month cost you almost nothing and the migration is just re-entering API keys in Open WebUI.