WorkBuro：一个快捷键唤起的 OpenAI 兼容桌面聊天客户端

Press a key, ask, and get an answer you can use. A small desktop client that talks to any OpenAI-compatible endpoint, including a model running on your own machine.

What it is · Why we made it · What it does · Main features · Install · First run · Keyboard

What it is

WorkBuro is a lightweight, powerful chat app for Windows and macOS that talks to any OpenAI-compatible LLM, answers on one keystroke, can run a model on your own machine with no account at all, and turns any of them into an agent when a job needs hands.

Why we made it

• The best part of an AI assistant is the model, and the worst part is everything in front of it: the tab, the sign-in, the wait. A powerful model deserves a lightweight window, not a browser tab and a queue.
• A hotkey beats a tab. Alt+Space over whatever you are doing, a question, an answer, gone. Lightweight enough to live on a keypress, and light enough to leave running all day.
• We wanted quick access to agents inside the same chat app. Flip Agent on and the model gets hands in a folder you can see, then flip it off again. No second app, no terminal, no session left running.
• It should do something, not only talk: ask for a spreadsheet and get a real .xlsx in a real folder.
• We wanted no second subscription. Point it at the endpoint you already pay for, or at a model on your own machine.
• The interface is familiar, and everything underneath is yours: your key, your endpoint, your files.
• It should be honest. If something does not work, it says so rather than quietly failing.

What it does

• It is summoned, not opened. A global hotkey (Alt+Space by default, rebindable) brings a lightweight window over your work and puts it away again. A tray icon means a hidden window is never a lost one.

• Answers you can use. Streamed markdown, tables, syntax-highlighted code with a copy button, and Stop, Regenerate and copy on every reply. It looks like the powerful chat client you already know, and it is small enough to forget about.

• Real files out. Say "save a CSV of three fruits and their quantities" and you get a real .csv, .xlsx, .docx, .pdf, .md, .json or .txt in Documents\WorkBuro, shown on the reply with Open and Show in folder.

• Documents in. Attach a PDF, Word, Excel, CSV, Markdown or text file and the model answers from what is inside it. One it cannot read comes back as a reason you can see, not an empty answer.

• Pictures both ways. Ask to see something and it draws it. Send a picture and it can change it from your words, or look at it and answer.

  

• A brief burst of agent work. Flip Agent on and the model gets shell and file tools inside a folder you can see, then you flip it off again. Nothing installs or schedules in the background. An attached picture is read as a picture rather than guessed at from its name, and switching Agent on part way through a chat picks up the conversation you were already having instead of starting from scratch.

• No key at all? It will run a lightweight model on your machine. One 429 MB download, then about a second to an answer, with no account and no browser.

The agent at work in a small window: it lists the folder, reads a file, writes one, and hands you the finished file with Open next to it. Same window, same keystroke, no second app.

Main features

• Any OpenAI-compatible endpoint. OpenCode Zen, OpenAI, OpenRouter, Groq, Together, DeepSeek, Mistral, LM Studio, Ollama, vLLM, llama.cpp, LocalAI, or your own gateway. One base URL field, with presets for the common ones.

• Model discovery and real capability checks. It reads /models from your endpoint, and the lightning button makes a real call to find out whether a model takes images, whether it reasons, and which wire protocol it wants. What it learns is remembered, so you stop guessing which of three similarly named models can really see.

• It copes with awkward endpoints. Chat Completions first, and the Responses API by itself when an endpoint refuses the first. Relay quirks such as session headers, a missing User-Agent, or /v1 pasted twice become readable errors instead of mystery HTML.

• Tools the model can call. Web search through your own SearXNG, read a page, weather, time, make a file, your own connected apps through your own Composio key, and any MCP servers you add.

  

• Images. fal.ai's live catalogue, each endpoint with its published price, prompts written by the model, progress while it draws, and a gallery in the sidebar of everything you have sent or drawn.

  

• No server to run. Nothing to install on a box, no container, no sign-in screen. If you have tried Open WebUI and wanted the same thing without running a server, this is that: an alternative to a ChatGPT subscription in the way that matters, because the key, the endpoint and the files stay yours.

• Made for a desktop. A small window that remembers its size, starts with Windows if you want, stays above other windows, and tucks itself away 30 seconds after you leave it, unless an answer is still arriving.

• Local-first and quiet. Lightweight where it counts: no telemetry and no backend. Every conversation lives in one JSON file you can open, back up or hand-edit.

• Small enough to read. A few thousand lines of TypeScript and CommonJS, powerful without being heavy, and a self-test that drives the real window instead of a mock.

Common questions

Is this a ChatGPT alternative? In the way people usually mean it, yes. It is a chat app that works like ChatGPT: streaming answers, a composer, a sidebar, an image mode, tools and agents. What is different is the model behind it, which is whatever you point it at, not one company's model behind one company's account. It is not affiliated with OpenAI and it does not claim to be ChatGPT.

Is there a free option? The app costs nothing and there is no subscription. With no API key at all it will download a lightweight model (429 MB) and run it on your machine, which is free to use from then on. If you already pay for a model, it uses your own key instead, and either way there is no second bill.

Why not just use Open WebUI? Open WebUI is a web app you install, run and sign into, usually in Docker and usually on a server you maintain. WorkBuro is a desktop app for Windows and macOS: install it, press Alt+Space, get an answer, close it. No server, no browser, no account. If you want a shared web UI for a team, use Open WebUI. If you want a desktop client that writes a real file into a real folder, this is the one.

Do I need an account? No. There is no sign-in, no telemetry, and no backend to sign into.

Does it only work with one model? No, and that is the point. OpenAI, OpenRouter, Groq, DeepSeek, Mistral, OpenCode Zen, LM Studio, Ollama, llama.cpp, vLLM, LocalAI, or your own gateway. If it speaks the OpenAI-compatible API, it works.

Install

Download from Releases:

Neither build is signed, so the system will ask the first time. On Windows, SmartScreen says "unknown publisher": choose More info, then Run anyway. On macOS, Gatekeeper says the app cannot be checked: right click the app, choose Open, then choose Open again. A SHA256SUMS-x.y.z.txt for Windows and a SHA256SUMS-x.y.z-mac.txt for macOS sit on each release if you want to check what you downloaded.

Or build it yourself:

git clone git@github.com:mesfeir/workburo.git
cd workburo
npm install
npm run dist     # -> release/WorkBuro-Setup-<version>.exe, or a .dmg on macOS

npm run dist makes an NSIS installer and a portable .exe on Windows, and a disk image and a zipped app on macOS.

First run

1. Open Settings (the gear, bottom-left, or Ctrl+,) and go to the connection tab.
2. No key yet? The first thing on that tab offers to run a small model on this machine. 429 MB, one download, and it answers from then on with no account.
3. Have a key? Pick a preset, or paste a base URL such as https://opencode.ai/zen/go/v1, paste your key, and press Test connection. The model list loads on success.
4. In Models, press the lightning button on the models you care about so their real abilities are written down.
5. For pictures, paste a fal.ai key in Settings → Images. That is the whole setup.

Keyboard

If Alt+Space is already taken, which happens on plenty of machines, WorkBuro walks a fallback list and shows an amber banner naming the key that is live. It never fails silently, and it never leaves you unable to reach the window.

Privacy

• No telemetry, no analytics, no backend. There is nothing to phone home to.
• Your key stays on your machine. It is held in the main process and leaves only as an Authorization header to the endpoint you configured. It is never exposed to the page.
• Prompts go to your endpoint, plus your search endpoint and Open-Meteo if you switch those tools on. Nothing else is contacted.
• One honest caveat. The key is stored in plain text in the app's own JSON file, at %APPDATA%\WorkBuro\zen-chat-store.json on Windows and ~/Library/Application Support/WorkBuro/zen-chat-store.json on macOS. That is a trade for transparency and hand-editability, and it is worth knowing before you file a security issue. Encrypting it with the system keychain is on the list.

Build from source

npm install
npm run dev      # Vite dev server plus Electron with devtools
npm run build    # type-check and bundle the renderer into dist/
npm run dist     # Windows installer and portable exe, or a macOS disk image, into release/
npm test         # the offline suites, including the local model

There is also an end-to-end harness that drives the real window: typing into the real composer, attaching a real document, switching models, reloading the app, and asserting on what actually rendered. The key goes in the environment, never on a command line where another process could read it.

Known limitations

• Neither build is signed. SmartScreen warns on Windows and Gatekeeper asks on macOS until that changes, which needs a paid certificate on one and an Apple Developer account on the other.
• macOS is Apple silicon only. There is no Intel build, and the macOS build is packaged and tested on one machine rather than across a range of Macs.
• Image generation is hosted. It needs a fal.ai key, and your prompts leave the machine when you use it. Local generation was measured on real hardware and the numbers are in docs/image-generation.md, but it is not wired in yet.
• MCP servers are stdio only. A server is a program run on this machine. Servers reachable only over HTTP are not supported yet.
• A very small model is a demo, not a worker. The 429 MB model that answers in about a second is good for a question and a reply. It is not the model to hand shell tools to.

Contributing

Issues and pull requests are welcome. Two rules this project holds itself to:

1. Never fake a capability. If it does not work yet, disable it and label it. Never stub something so it looks finished.
2. Never silently degrade. If a fallback kicks in, say so in the UI.

Run npm run build, which type-checks, before opening a pull request. And never commit a key: scripts/check-secrets.cjs scans the tree for key-shaped literals and can install itself as a pre-commit hook.

Licence

Not chosen yet. No LICENSE file has been added, so all rights are reserved for now.