中文
Home / NextChat Deployment Guide: One-Click AI Chat Assistant on Vercel

NextChat Deployment Guide: One-Click AI Chat Assistant on Vercel

Xiaozha 4 tags
Glowing AI chat interface on screen

You use ChatGPT, Claude, and Gemini — but each has its own website, its own login, and its own $20/month subscription. Switching between them is a chore, and everything you type lives on someone else’s server.

NextChat (formerly ChatGPT-Next-Web) fixes all of that. It is a lightweight, open-source AI chat client that unifies 16+ major LLM providers — OpenAI, Claude, DeepSeek, Gemini, and more — behind a single, clean interface. Your conversations are stored locally in your browser, your API keys stay in your own deployment, and the whole thing runs for free on Vercel’s free tier.

This complete deployment guide covers what NextChat is, its core features, one-click Vercel deployment, Docker deployment to your own server, every important environment variable, connecting DeepSeek/Claude/Gemini/SiliconFlow, custom domains with Cloudflare, data backup, and the most common pitfalls.

What Is NextChat and Why Self-Deploy It?

NextChat is an open-source web client for large language models. Instead of a “one provider” interface, it is a universal remote control for AI: pick a model from a dropdown and chat — same UI, same history, zero learning curve.

ConcernOfficial ChatGPTNextChat
Monthly fee$20/monthFree (you only pay API usage)
Deployment difficultyNone (hosted)~1 minute on Vercel
Model selectionGPT series only16+ providers, freely switchable
Data privacyCloud storageLocal browser storage
Custom domainNot supportedSupported
API keysHandled by OpenAIYou control them

Why self-deploy?

  • Privacy — conversations live in your browser’s localStorage, not in a third-party cloud.
  • Cost — you pay only for actual API usage; a personal DeepSeek or Gemini setup costs pennies a month.
  • Freedom — switch models mid-conversation and use the same client for OpenAI, Claude, DeepSeek, and Gemini.
  • No subscriptions — never pay a flat monthly fee for chat again.

Core Features

NextChat is deceptively full-featured. The features you will actually use:

FeatureWhat it does
16+ model providersOpenAI, Claude, DeepSeek, Gemini, and domestic Chinese models (Ernie, Qwen, Hunyuan, Doubao)
Local storageAll chats and settings saved in localStorage; nothing leaves your browser except API calls
Markdown & code renderingFull Markdown, syntax-highlighted code blocks, LaTeX
Masks (prompt presets)Dozens of role templates — copywriter, code reviewer, translator — out of the box
Artifacts previewLive preview of generated HTML/web pages inside the chat
Voice inputSpeak instead of type with voice-capable models
MCP supportModel Context Protocol plugins: web search, calculators, custom APIs
SyncOptional WebDAV / Upstash backup so history survives browser wipes
Multi-platformWeb, PWA (installable), and a tiny desktop client

Method 1: One-Click Deploy on Vercel

This is the fastest path — no server, no Docker, no terminal required.

Step 1 — Fork the repository

Open the official repo and click Fork in the top-right corner:

Wait a few seconds for the fork to complete.

Step 2 — Deploy on Vercel

  1. Go to vercel.com/new and import your fork.
  2. Vercel auto-detects the project; leave the build settings at their defaults.
  3. Before clicking Deploy, add the two key environment variables:
VariableValuePurpose
OPENAI_API_KEYsk-...Your API key (OpenAI, DeepSeek, or any compatible provider)
CODEe.g. my-secret-passwordAccess password so strangers can’t burn your API credits
  1. Click Deploy. After ~5 minutes the status flips to Deployed.

Step 3 — Open your assistant

Vercel assigns you a URL like https://nextchat-yourname.vercel.app. Open it, enter the CODE password you set, and you’re chatting.

Note: If you skip CODE, the deployment is publicly accessible and anyone who finds the URL can use your API credits. Always set a password, even for personal use.

Method 2: Docker Deployment to Your Own Server

Prefer full control? Run NextChat on your own machine or VPS with a single docker run:

docker run -d --name nextchat \
  -p 3000:3000 \
  -e OPENAI_API_KEY=sk-your-key \
  -e CODE=my-secret-password \
  -e BASE_URL=https://api.deepseek.com \
  --restart unless-stopped \
  yidadaa/chatgpt-next-web

Or with Docker Compose — cleaner and easier to update:

services:
  nextchat:
    image: yidadaa/chatgpt-next-web
    container_name: nextchat
    restart: unless-stopped
    ports:
      - "3000:3000"
    environment:
      OPENAI_API_KEY: sk-your-key
      CODE: my-secret-password
      BASE_URL: https://api.deepseek.com
docker compose up -d

Open http://localhost:3000 (or http://server-ip:3000). To update:

docker compose pull
docker compose up -d

The image is also published as ghcr.io/chatgpt-next-web/chatgpt-next-web if you prefer GitHub Container Registry.

Environment Variables You Should Know

Beyond the two basics, these variables give you real control:

VariablePurposeExample
OPENAI_API_KEYDefault API keysk-...
CODEAccess password (comma-separated for multiple users)pass1,pass2
BASE_URLAPI base URL (defaults to https://api.openai.com/v1)https://api.deepseek.com
ANTHROPIC_API_KEYClaude’s keysk-ant-...
ANTHROPIC_BASE_URLClaude API endpoint overridehttps://api.anthropic.com
OPENAI_MODEL_LISTVisible model list (defaults to all)-all,+gpt-4o,+gpt-4o-mini
CUSTOM_MODELSAdd models beyond the default list+deepseek-chat=>DeepSeek V3
DEFAULT_MODELModel preselected at startupgpt-4o
ENABLE_MCPTurn on MCP plugin supporttrue

The model-list syntax deserves attention. -all,+gpt-4o means “hide everything except gpt-4o”. A common pattern is to expose only the models you actually have access to:

OPENAI_MODEL_LIST=-all,+gpt-4o,+gpt-4o-mini,+deepseek-chat,+deepseek-reasoner

Connecting DeepSeek

DeepSeek’s API is OpenAI-compatible, which makes it a drop-in replacement. This is the most popular integration because DeepSeek is dramatically cheaper than OpenAI while being excellent at coding and reasoning.

Set these environment variables (Vercel: Project → Settings → Environment Variables; Docker: the environment: block):

OPENAI_API_KEY=sk-deepseek-your-key
BASE_URL=https://api.deepseek.com
OPENAI_MODEL_LIST=-all,+deepseek-chat,+deepseek-reasoner

Then, in the chat UI (bottom-right settings):

  1. Scroll to Custom Interface and enable it.
  2. AI provider: DeepSeek (or keep OpenAI-compatible with the BASE_URL above).
  3. API key: your DeepSeek key.
  4. Model: deepseek-chat (fast, general) or deepseek-reasoner (reasoning).
  5. Click back and send a message — replies now come from DeepSeek.

Why DeepSeek is a great default

  • Pricing — a fraction of OpenAI’s per-token rates.
  • OpenAI-compatible API — zero code changes, just a different BASE_URL and key.
  • Free credits on signup — new accounts get promotional credit, so you can test the whole setup for free.

Connecting Claude and Gemini

NextChat treats each provider as a selectable option in the UI, but you can also wire the keys through environment variables.

Claude (Anthropic)

ANTHROPIC_API_KEY=sk-ant-xxxxxxxx
ANTHROPIC_BASE_URL=https://api.anthropic.com
OPENAI_MODEL_LIST=-all,+claude-3-7-sonnet,+claude-sonnet-4

Or configure entirely in the UI: settings → choose Claude as the provider, paste the key, pick a model like claude-sonnet-4.

Gemini (Google)

Gemini’s API also speaks the OpenAI protocol through a compatibility endpoint, so you can point NextChat at it directly:

OPENAI_API_KEY=AIza-your-gemini-key
BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai
OPENAI_MODEL_LIST=-all,+gemini-2.0-flash,+gemini-2.5-pro

Gemini’s free tier is generous — for many users it becomes the default model.

Model switching in practice

Once multiple providers are configured, the model dropdown in the top-left lists everything. You can switch mid-conversation: the same chat thread continues with a different underlying model.

Connecting a Domestic Relay API (SiliconFlow)

For users who need fast access in China — or simply want cheap, aggregated access to many models — relay services are the answer. SiliconFlow (硅基流动) is a popular one: it hosts open models (Qwen, DeepSeek, GLM, Llama) behind an OpenAI-compatible API and offers free starter credits.

Set:

OPENAI_API_KEY=sk-siliconflow-your-key
BASE_URL=https://api.siliconflow.cn/v1
OPENAI_MODEL_LIST=-all,+Qwen/Qwen2.5-72B-Instruct,+deepseek-ai/DeepSeek-V3
  1. Register at SiliconFlow, grab an API key from the console.
  2. Pick a model slug from their model list (e.g. Qwen/Qwen2.5-72B-Instruct).
  3. Paste key + base URL into NextChat (env vars or the UI).
  4. Model names in OPENAI_MODEL_LIST must match the provider’s exact slugs.

The same pattern works with any other OpenAI-compatible relay — shaw-openai-style free relays, corporate gateways, or self-hosted gateways like one-api/new-api. Just change BASE_URL and the key.

Custom Domain and Cloudflare Proxy

Vercel’s *.vercel.app URL works, but a custom domain looks professional and is more reliable.

Bind a domain in Vercel

  1. Open the project → Settings → Domains.
  2. Enter your domain, e.g. ai.example.com.
  3. Vercel shows you a DNS record to add (usually a CNAME or A record).
  4. If the domain is on Cloudflare, Vercel’s auto-config flow can set it up for you after authorization.

Using Cloudflare DNS in front of Vercel gives you caching, DDoS protection, and your whole zone in one place:

ai.example.com  CNAME  cname.vercel-dns.com   (Proxied, orange cloud)

Once DNS propagates and the certificate issues, https://ai.example.com opens your assistant.

For China access: vercel.app is blocked in some regions. A custom domain via Cloudflare (or a relay/edge proxy) is the standard workaround. Note that for Chinese mainland endpoints, a domain on a mainland-CDN-friendly provider or a domestic relay backend is often smoother than going through Cloudflare’s overseas edge.

Data Backup: Exporting Your localStorage

NextChat stores everything — conversations, settings, masks — in your browser’s localStorage. That means clearing your browser data wipes your history. Back it up.

Manual export

  1. Open Settings → Data.
  2. Click Export All Data → download the JSON file.
  3. Keep it somewhere safe (a password manager, encrypted drive, or your NAS).
  4. To restore, Import Data and select the file.

Automatic sync

Enable WebDAV or Upstash sync in Settings → Data:

  • WebDAV — point it at any WebDAV server (Nextcloud, Synology) for seamless cross-device sync.
  • Upstash Redis — a free hosted Redis for syncing chat state across devices.

With either enabled, you can wipe your browser, log back in, and your full history returns.

Common Problems and Fixes

Q1: CORS errors when calling the API

If you use a relay or a non-official BASE_URL, the browser may block the API call with a CORS error. Fixes, in order:

  1. Switch the provider to run through NextChat’s server-side proxy (env-var config rather than browser-only settings).
  2. Ensure the relay explicitly allows your origin (most OpenAI-compatible relays do).
  3. As a last resort, proxy through your own reverse proxy that adds CORS headers.

Q2: Forgot the access password (CODE)

  • Vercel: Settings → Environment Variables → update CODE → Redeploy (a new deployment applies env changes).
  • Docker: edit the environment: block, then docker compose up -d.
  • The password applies at login; changing it kicks everyone and forces a fresh login.

Q3: Models don’t show in the dropdown

The model list is controlled by OPENAI_MODEL_LIST / CUSTOM_MODELS. If you set -all, every model is hidden and only the ones you explicitly +-prefixed appear. Fix: list exactly the models you have access to.

Q4: “Error: OPENAI_API_KEY is not set”

The key is missing or empty in your deployment env. Also confirm BASE_URL matches the provider — many relays use a different path (/v1 suffix matters).

Q5: Messages fail after a while

Usually one of: API credits exhausted (check your provider dashboard), a free-tier rate limit, or a relay outage. Try switching to a different model or provider to isolate the cause.

Summary

NextChat is the fastest way to own a private, multi-model AI chat assistant at zero deployment cost. In this guide you learned how to:

  1. Understand what NextChat is — an open-source client unifying 16+ LLM providers.
  2. Use the core features — local storage, Markdown, Masks, Artifacts, MCP.
  3. Deploy on Vercel in one click — fork, import, add OPENAI_API_KEY + CODE.
  4. Deploy with Docker — docker run or Compose on your own server.
  5. Configure environment variables — OPENAI_API_KEY, CODE, BASE_URL, model lists.
  6. Connect DeepSeek — an OpenAI-compatible, budget-friendly default.
  7. Connect Claude and Gemini — including Gemini’s free tier.
  8. Use domestic relays like SiliconFlow — cheap aggregated model access.
  9. Bind a custom domain and front it with Cloudflare.
  10. Back up chat data — manual export and WebDAV/Upstash sync.
  11. Fix common issues — CORS, passwords, model lists, missing keys.

The quickest way to try it: fork the repo, deploy on Vercel with OPENAI_API_KEY and CODE, and start a conversation with DeepSeek within ten minutes. If you found this useful, you might also enjoy our guides on running LLMs locally with Ollama and exposing local AI services with Cloudflare Tunnel. Happy chatting! 🤖

AITutorialFree ToolsLLM