An offline-first AI writing assistant that lives in your system tray. Hit a global hotkey, send your selection or clipboard to Claude, GPT, Gemini (or 11 other providers), and paste the result back — all without leaving the app you're in.
Bring your own API keys — none of your prompts, selections, or clipboard contents leave your machine except the request you explicitly send to your chosen provider. Chat history is stored locally in IndexedDB.
Most AI front-ends are web tabs. NeuroWriter is a desktop launcher:
- Global hotkey — summon a chat window from anywhere without alt-tabbing.
- Clipboard pipeline — the last thing you copied is one keystroke away from "rewrite this", "translate this", "summarize this".
- System tray — quick actions live in your menu bar, not in a browser tab.
- 14+ providers, one UI — switch models per-conversation. Your key, your rate limit, your bill.
- Offline-first — settings, conversations, masks, and prompts are stored locally. No accounts, no cloud sync.
yarn install
yarn dev
# open http://localhost:3000yarn install
yarn app:devyarn app:build
# Outputs in src-tauri/target/release/bundle/Drop an API key into Settings (or .env.local) and the provider lights up.
No SDK, no account creation, no telemetry — every request goes browser →
provider directly.
| Provider | Models exposed | Key var |
|---|---|---|
| OpenAI | GPT-4o, GPT-4, GPT-3.5, GPT-5 | OPENAI_API_KEY |
| Anthropic | Claude Sonnet, Claude Opus, Claude Haiku | ANTHROPIC_API_KEY |
| Gemini 1.5 Pro, Gemini Flash | GOOGLE_API_KEY |
|
| DeepSeek | DeepSeek-V3, DeepSeek-Coder | DEEPSEEK_API_KEY |
| xAI | Grok | XAI_API_KEY |
| Moonshot | Kimi (long-context) | MOONSHOT_API_KEY |
| Alibaba | Qwen | ALIBABA_API_KEY |
| Baidu | Ernie | BAIDU_API_KEY |
| ByteDance | Doubao | BYTEDANCE_API_KEY |
| Tencent | Hunyuan | TENCENT_API_KEY |
| Zhipu | GLM | CHATGLM_API_KEY |
| iFlytek | Spark | IFLYTEK_API_KEY |
| SiliconFlow | Aggregator (Qwen / Llama / Mistral mirrors) | SILICONFLOW_API_KEY |
| 302.AI | Aggregator (pay-as-you-go to anything) | AI302_API_KEY |
Add a provider that's not listed via
CUSTOM_MODELS=+provider/model-name. Hide ones you don't want with-provider/model-name.
- Chat with code blocks — syntax highlighting, copy buttons, math via KaTeX, diagrams via Mermaid.
- Masks — reusable system prompts ("Senior code reviewer", "ESL editor", "Stand-up summarizer") that you can pin or share via URL.
- Realtime voice — Azure realtime-audio for streamed voice chat.
- MCP tools — set
ENABLE_MCP=trueto expose Model Context Protocol servers as tools your chat can call. - Image input — paste / drag images into a chat; HEIC is converted in
the browser via
heic2any. - PWA — install from the browser if you don't want the desktop app.
- Export anything — Markdown, PNG, JSON. No vendor lock-in on your conversation history.
Placeholder — drop screenshots into
docs/screenshots/and they'll render here. Filename suggestions:tray-launcher.png,chat-window.png,mask-picker.png,settings.png.
NeuroWriter reads from environment variables (server-side) and Settings (client-side, persisted to IndexedDB). The web build expects env vars; the desktop build can use either.
Common variables (see .env.template for the full list):
| Var | Purpose |
|---|---|
OPENAI_API_KEY |
Server-side OpenAI key. Required for SaaS deploys. |
ANTHROPIC_API_KEY |
Server-side Anthropic key. |
GOOGLE_API_KEY |
Server-side Gemini key. |
CODE |
Comma-separated access passwords for self-hosted. |
PROXY_URL |
Outbound HTTP proxy (e.g. for region-locked APIs). |
CUSTOM_MODELS |
Whitelist / blacklist / rename models. |
DEFAULT_MODEL |
Model to select on new conversations. |
HIDE_USER_API_KEY |
Prevent users from supplying their own key. |
ENABLE_MCP |
Enable Model Context Protocol tools (true). |
For desktop, users typically leave server-side vars empty and enter their own keys in Settings — that's the point of BYO.
| Layer | Tech |
|---|---|
| Web framework | Next.js 14 (App Router off — Pages for static export) |
| Desktop shell | Tauri 1.5 (Rust) |
| UI | React 18 + Sass |
| State | Zustand |
| Persistence | IndexedDB (via idb-keyval) |
| Markdown | react-markdown + remark-gfm + rehype-katex |
| Diagrams | Mermaid 10 |
| Search | Fuse.js |
| Voice | rt-client (Azure Realtime Audio) |
| Tools | @modelcontextprotocol/sdk (optional, behind ENABLE_MCP) |
yarn dev # Web dev server (http://localhost:3000)
yarn build # Standalone Next.js build (for self-host / Docker)
yarn export # Static export — feeds the Tauri bundle
yarn app:dev # Tauri desktop in dev mode
yarn app:build # Tauri installers (.exe / .msi / .dmg / .deb / .AppImage)
yarn lint # ESLint
yarn test # Jest in watch mode
yarn test:ci # Jest single-run for CIapp/ Next.js pages, components, stores, providers
client/platforms/ One file per AI provider (openai, anthropic, …)
masks/ Mask definitions + build script
components/ React components (chat, settings, sidebar, mask picker)
src-tauri/ Rust shell, system tray, global hotkey
public/ Static assets, icons, service worker, prompts.json
docs/ Documentation (screenshots go in docs/screenshots/)
scripts/ Helper scripts (proxy init, prompt fetcher)
This is a vanilla Next.js app — deploy with the platform's GitHub
integration. Set the env vars from .env.template, push to main.
docker build -t neurowriter .
docker run -d -p 3000:3000 \
-e OPENAI_API_KEY=sk-... \
-e CODE=your-access-password \
--name neurowriter \
neurowriterOr docker compose up -d.
yarn install
yarn build
yarn start
# Listens on :3000. Front with nginx for TLS.- No accounts. No sign-up flow, no email collection, no cloud sync.
- No proxying by default. Your API calls go browser → provider directly.
The Next.js server is only a proxy when you set
BASE_URLor run the SaaS build with a server-side key. - No analytics in the desktop build. The web
@vercel/analyticsimport is only active when deployed on Vercel. - History is local. Conversations, masks, settings, and access codes
live in your browser's IndexedDB (web) or
~/.neurowriter(desktop). Clear browser data → conversations gone.
- Tauri 1.5, not 2. The desktop build is on Tauri v1; mobile (iOS /
Android) targets aren't wired up. The upstream
ChatGPTNextWeb/NextChatships iOS as a separate native app. - Updater endpoint in
src-tauri/tauri.conf.jsonstill points at the upstream release feed. Repoint it before publishing your own releases or the desktop app will offer upstream updates. - Server-side proxy mode assumes one user per deployment (or one shared
CODE). For real multi-tenant use, swap in your own auth.
NeuroWriter is a fork of ChatGPTNextWeb/NextChat (originally
Yidadaa/ChatGPT-Next-Web), with the following changes layered on top:
- Rebranded as NeuroWriter with desktop-launcher positioning.
- Added global hotkey, system tray integration, and a clipboard pipeline.
- Removed SaaS / Enterprise / sponsor banners from the UI and README.
All credit for the original chat client, provider integrations, mask system, and Next.js + Tauri scaffolding goes to the upstream maintainers and contributors. MIT licensed.
MIT — same as upstream.