WebBrain Companion

The AI plans you already pay for, ready for WebBrain

WebBrain Companion lives in your menu bar or system tray. Sign in to Claude, ChatGPT, Google, Grok or Kimi once, and WebBrain reaches all of them at one address on your computer. The tray shows how much of each plan is left and when it resets.

Download Version 0.1.2 for macOS, Windows and Linux. Free and open source.
Menu on

Everything else happens in one small panel

Click the icon to open it. One page holds the address for WebBrain, every plan with what's left, and the two local models; click a row for its details, or the gear for settings. It follows your light or dark setting and speaks English and Turkish.

  • The panel: the address http://127.0.0.1:8327/v1 with a Copy button, usage meters for Claude Max, ChatGPT Pro, Google AI Pro and SuperGrok, and Bonsai 2 27B running next to the d1 decision model.

    One page

    The address WebBrain needs, each plan with the limit closest to running out, and the local models with their main button. No API key on this computer.

  • The add subscription sheet listing Claude, ChatGPT, Google, Grok and Kimi, each with how you sign in.

    Add a subscription

    Sign in on the provider's own page. The login stays on your computer.

  • The Bonsai 2 27B details over the panel: running on the Apple GPU with image input on, 65,536 tokens of context, and its options.

    Local models

    Click Bonsai 2 27B to see where it runs and whether images fit, and to change its options. d1 opens the same way.

Bonsai 2 27B, sized to your machine

PrismML's Ternary Bonsai 2 27B packs 27 billion parameters into 5.95 GB. The companion installs it with its Q8_0 vision projector and PrismML's build of llama.cpp, or as the MLX pack on Apple Silicon, and lists it next to your subscriptions as bonsai-2-27b.

Before anything downloads, it works out where the model runs and whether image input fits. The vision projector goes on the GPU when about 1 GB of GPU memory is left after the model and its context, on the CPU when only system RAM has room, and stays off otherwise. You can override the choice in the panel.

What the companion picks on typical machines
MachineRuns onImage inputContext
MacBook Air, 8 GBApple GPU with Metal. It needs more memory than the Mac has, so expect swapping.Off8K tokens
MacBook Pro, 16 GBApple GPU with MetalOn, GPU16K tokens
Laptop with an RTX 4060 (8 GB) and 16 GB RAMCUDA, 57 of 64 layers on the GPUOn, CPU8K tokens
Desktop with an RTX 4090 (24 GB) and 64 GB RAMCUDA, all on the GPUOn, GPU64K tokens
Desktop with a Radeon RX 7600 (8 GB) and 32 GB RAMVulkan, 61 of 64 layers on the GPUOn, CPU8K tokens
PC without a graphics card, 16 GB RAMCPUOn, CPU8K tokens
PC without a graphics card, 8 GB RAMCPU, with little room to spareOff8K tokens

Worked out by the companion's planner from total and free memory; the numbers on your machine at install time decide. On Apple Silicon the MLX pack carries its own vision tower, so images are always on.

WebBrain's decision model, outside the browser

WebBrain's d1 model answers yes/no, choice and score questions about a page and up to four screenshots, such as whether a task is finished. The companion runs it on ONNX Runtime and answers over HTTP, so a browser tab doesn't have to load 1.9 GB. Its Go port returns the same tokens and probabilities as WebBrain's JavaScript adapter. Native d1 is available on Windows, Linux and Apple Silicon Macs; it is unavailable on Intel Macs.

Ask

curl http://127.0.0.1:8327/companion/v1/decide \
  -H "Content-Type: application/json" \
  -d '{"state": "Order #1042 is confirmed.",
       "questions": {"completion": {
         "type": "choice",
         "instructions": "Did the task finish?",
         "criteria": {"completed": "Confirmed",
                      "not_completed": "Failed or unfinished",
                      "unknown": "Not established"}}}}'

Answer (273 ms on the CPU runtime)

{"answers": {"completion": {
   "choice": "completed",
   "confidence": 0.99997,
   "probabilities": {"completed": 0.99997,
                     "not_completed": 0.000002,
                     "unknown": 0.000033}}},
 "diagnostics": {"runtime": "onnxruntime 1.31.0 (CPU)", …}}

Connect WebBrain

You do this once. Apps on this computer need no API key.

  1. Install the companion and open it. Add a subscription in its panel and finish the sign-in in your browser.
  2. The default address is http://127.0.0.1:8327/v1. If the port changed, copy the address shown in the panel or tray menu.
  3. In WebBrain, open the Providers page (More… in the side panel's model menu) and choose WebBrain Companion. Keep its default address unless the app shows another port. No API key is needed on this computer.
  4. Click Load Models, pick a model, Save, and run Test Connection before activating the card. Turn on vision only for a model that accepts images.

Any OpenAI, Anthropic or Gemini client on this computer can use the same address. To use it from another device, turn on network access in the panel's settings; that device then needs the API key the panel shows.

Download

Version 0.1.2. Each package holds a single program; your settings and logins live in your user folder.

macOS

macOS 12 or later, Apple Silicon and Intel (d1 requires Apple Silicon)

Your system

Move WebBrain Companion to Applications. The build is not notarized by Apple: the first time, allow it in System Settings, Privacy & Security (Open Anyway). On macOS 14 and earlier, Control-click it and choose Open.

Windows

Windows 10 and 11, x64 and Arm64

Your system

Run the installer: it installs for your account, without an administrator prompt, and adds WebBrain Companion to the Start menu. Until the installer is code-signed, SmartScreen may say "Windows protected your PC": choose More info, then Run anyway.

Linux

x64 and Arm64

Your system

Unpack and run webbrain-companion. The icon needs a StatusNotifierItem tray: KDE, XFCE, or GNOME with the AppIndicator extension.

SHA-256 checksumsSource codeAll releases

Before you sign in

What stays on your computer

  • On this computer no API key is needed: the companion answers only requests from 127.0.0.1 and turns away web pages, so a site you visit can't spend your plans. Browser extensions such as WebBrain and apps on this computer can use it.
  • Other devices can connect only if you turn on network access in the panel's settings. It then listens on 0.0.0.0, and they need the API key the panel shows.
  • Logins are stored in your user folder and sent only to the provider they belong to.
  • There are no analytics. Apart from the providers you sign in to, it only fetches CLIProxyAPI's public model list, and, when you install a model, its files from Hugging Face, GitHub and, for MLX, PyPI.
  • Prompts sent to a subscription still go to that provider. Only Bonsai 2 27B and d1 run on your machine.

What your provider decides

  • A subscription that works through a proxy is not the same as one the provider allows. Anthropic's consumer terms restrict automated access outside an API key, and Google's Gemini CLI terms say its sign-in may not be used with third-party tools.
  • Check the terms of your plan and region before you add it. If an account matters to you, a provider API key is the safer route.
  • Keep network access off on public Wi-Fi, and never forward port 8327 to the internet. The API key is not internet-grade security.

Connect WebBrain to Companion