_{The Windows client. One message gets the Bot to look something up, cite its sources, and draw a picture on the side. The model, web access and image generation all go through your own APIs.}

_{The screenshots show the Chinese UI; the apps also support English.}

ProkBot Plus is a modified version of the open-source project ProkBot. The idea behind ProkBot: the original Grok Bot apps and Agent work well, but they can only connect to the official cloud. ProkBot runs the server side (the Host and the Linux desktop used for execution) in Docker on your own machine, then patches the original apps so they connect to your server.

Plus adds three things on top of that: a Windows client; two more model APIs, Anthropic Messages and OpenAI Responses (upstream supports only OpenAI Chat); and a set of fixes that make web search, image generation and local commands work smoothly in real use.

This is an unofficial project. It has no connection to xAI or the official Grok Bot. Noncommercial use only. See License and notices.

What it does

• Use the original Grok Bot. The interface, Agent, tool calls, memory, scheduled tasks and action approvals are all the original ones. Only the server side is now in your hands.
• Connect any model that supports tool calls. Claude, GPT, DeepSeek, Grok, or any aggregator gateway. Pick any of the three API formats, and set thinking effort anywhere from none to max. You can also let the Jev classifier pick the model by task difficulty.
• Let the Bot work on its own computer. The server comes with a Linux desktop. The Bot runs commands, opens a browser and handles files on it.
• Let the Bot operate your computer. On Mac it can run commands and control the graphical interface. On Windows it can run local commands, and it asks you in a pop-up before each one.
• Web access and image generation. Search and read web pages with the model's built-in web tools. Generate images with an OpenAI-compatible image API, with support for set aspect ratios and editing from a reference image.
• Your data stays on your machine. Chats, files, memory and Bot data are all stored on your server. Telemetry and automatic updates are turned off.
• Use it when you're out. Direct LAN connection, built-in Tailcat networking (no public IP needed), or your own HTTPS domain.

Supported systems

Part	Supported	Notes
Server	Linux x86_64 (Docker)	Tested on Ubuntu 24.04 + Docker 29. A NAS that supports Docker also works. ARM servers are not supported, and neither is deploying directly on Windows
Windows client	Windows 10 / 11, x64	New in this version. Portable version and installer, no administrator rights needed
Mac client	Apple Silicon, Grok Bot 0.71.2	From upstream
iOS client	Grok Bot 1.16.0 (23152)	From upstream. You need your own decrypted IPA and signing certificate
Android, Linux desktop	Not supported	There is no official client for them

What's new in this version

Windows client

On first launch, enter the Host address and the 6-character pairing code.	When the Bot wants to run a command on your computer, it asks you first.

Upstream has only Mac and iOS. The Windows version is built by patching the official 0.71.2 Windows installer, and its interface matches the official Windows version:

• Stays in the system tray, system notifications, attachments, Tailcat;
• Local command execution (PowerShell, Chinese output is not garbled);
• Hands complex coding work to Codex on your computer;
• Builds both a portable version and an installer.

The patches are ported from the Mac version, and the structure of every segment is checked.

Three model APIs

Upstream supports only OpenAI Chat Completions. Plus adds Anthropic Messages and OpenAI Responses. All three formats are fully supported: images, files, tool calls, thinking effort, and reasoning carried across turns.

• Switching models or formats in the middle of a conversation doesn't cause errors.
• When a provider doesn't support an optional parameter, it is removed and the request is retried automatically.
• Automatic safety review also works with all three formats.

Web access and image generation

Upstream already connected the original web and image tools, but they had quite a few problems with newer models. Plus fixes these:

• Newer Claude models don't accept forced tool calls, so web access failed outright → it now switches to a regular tool call automatically;
• In a parallel search, if one search went over its usage limit, the whole search was marked as failed → now the result is used as long as one search succeeds;
• When the Bot asked to "draw a 16:9 image", the OpenAI image API returned an error → the ratio is now converted to the matching size automatically;
• Editing with a reference image returned an error → it now goes through the image edit endpoint.

Other fixes

• Message order. The original Bot splits a reply into several short messages. In the upstream implementation, these messages could arrive out of order, with the short ones first. Now they are delivered strictly in the order the model wrote them.
• Automatic review. When the model sometimes answered in the wrong format, the action used to be rejected right away, with nothing in the log. Now it retries once, and failures are logged.
• Linux deployment. Docker 29 uses the containerd image store by default, which broke the upstream build script. This is now supported.
• Command-line management. New host/tools/host-admin.mjs lets you set up the Host, configure models and create pairing codes without opening the web page. This makes things easier for scripts and AI.
• Docs. A new AI deployment guide that even an average model can follow.

See the CHANGELOG and the technical documentation for all changes.

Compared with the official Grok Bot

Benefits

• Use the models and providers you choose. Costs are billed through your own API, and you can also use relays or aggregator gateways.
• Chats, files and memory are all on your own server, with no telemetry.
• Works where the official service isn't available. You can also use it only on your LAN.
• Some things the official version doesn't have:
  • Edit the system prompt, and restore the original at any time;
  • Set the context length per model;
  • Usage statistics;
  • Tasks triggered by webhooks;
  • Write your own apps (plugins);
  • Bot collaboration across several Hosts;
  • Full-machine backup and restore.

Trade-offs

• You need a Linux server (at least 8 GB of memory and 30 GB of disk), and you handle upgrades and backups yourself.
• Versions are pinned: client 0.71.2, iOS 1.16.0, Host d0a4baa. When the official apps update, you have to wait for the patches to catch up before upgrading, and there are no automatic updates.
• Results depend on the model you choose. The official service is tuned for its own models. With other models, the experience may not be as good.
• Some official features are turned off or replaced:
  • Turned off: virtual cards, external message drafts, product feedback, passkey forwarding;
  • Replaced: voice, push notifications, image and video generation and others, which need providers you set up yourself;
  • Real-time calls are experimental for now.
• The clients are not code-signed: Windows shows a SmartScreen warning on first launch; on iOS you sign and install the app yourself.
• The Windows version can't yet let the Bot control the mouse and keyboard. The original component for this only exists for macOS.
• Modifying the official client may break the official terms of use. Use at your own risk.

What you need to deploy

Required	Requirements
Server	x86_64 Linux with 8 GB+ of memory and 30 GB+ of free disk, with Docker, Docker Compose and Python 3.10+ installed. The first build downloads about 8 GB
Model API	Any one of: Anthropic Messages, OpenAI Responses, OpenAI Chat Completions. It must support tool calls. For the Bot to operate a computer, it also needs image input
Computer to build the clients	Build the Windows client on Windows 10/11 x64 and the Mac client on an Apple Silicon Mac. Both need Node.js 22.12+. iOS needs a decrypted IPA and a signing certificate

Optional	Used for
A model with native web access	The Bot searches and reads web pages (for example Claude's web_search)
An OpenAI-compatible image API	The Bot generates and edits images
Bark or an Apple push certificate	Push notifications on phones
Speech recognition and synthesis services	Voice input, real-time calls
A domain and HTTPS	Access from outside without Tailcat

Install

The easiest way is to hand it to an AI that can run commands (Claude Code, Codex, Cursor and similar tools all work). Send it this:

Please deploy ProkBot Plus for me: https://github.com/ferretgeek/prokbot-plus
First read docs/AI-INSTALL.en.md in the repository in full, then follow its steps exactly.
Before you start, tell me what I need to prepare.

The AI first lists what you need to prepare. Then it deploys the server, configures the model, builds the clients step by step, and checks the result at the end.

If you'd rather do it yourself, follow the AI deployment guide. It is written for AI, but every step lists the commands and how to check them, so a person can follow it just as well.

Releases has the patch tools split into packages (server, Windows, Mac, iOS), so you can download only the part you need. Releases do not include patched clients: they contain the official Grok Bot program, so you have to generate them with the tools on your own machine.

Documentation

Document	Contents
AI deployment guide	Full steps for deploying from scratch. Every step says what you should see and what to do if it fails
Technical documentation	How it works, how each feature is implemented, what Plus changed, the layout, tests, extension points and known limitations
Windows client · Mac client · iOS client	How to use the patches for each platform
Split deployment, backups and app development	Advanced usage

Friends

• LINUX DO: a Chinese tech community where ProkBot Plus is discussed and released. Questions and ideas are welcome there.

License and notices

• This project is derived from ProkBot (based on its 2026-11-05 version, commit 3e9fe8c) and keeps its ProkBot Noncommercial License 1.0.
  • You may use, study, modify and distribute it for personal and noncommercial purposes;
  • You may not use it for any commercial purpose, including paid deployments, paid services and internal use in a company.
  • See NOTICE for the source and the list of changes.
• The repository contains no official Grok Bot programs. The patch tools download the official version on your own computer, verify it, and then patch it. The repository does not provide any original or modified IPA.
• Grok, Grok Bot and related trademarks belong to their respective owners. This project is not affiliated with or endorsed by them.
• Most of the code in this project was written with AI assistance. It is provided as is, without any warranty.
• To report a security issue, see SECURITY.md. To contribute, see CONTRIBUTING.md.