MagicHandy, an AI-enabled one-stop shop for strokers!

Software
1

MagicHandy is ready for early testing!

image
image2560×1274 218 KB

MagicHandy is the ground-up successor to StrokeGPT-ReVibed, rebuilt around a lighter Go core, a modern browser interface, and a more advanced motion architecture.

MagicHandy connects a local LLM to a Handy and lets it control motion using conversation, live feedback, patterns, and hands-free modes.

Put more bluntly: it is a local AI chatbot that fucks you.

The new app is rewritten in Go, making it much more lightweight, and every source of motion now goes through one shared engine. App stability and maintainability should be much improved.

Special thanks to roquebalaboa for contributing, and moving from working on LSO to helping with MagicHandy!

What MagicHandy Can Do

  • Chat-driven motion: Talk to a local model through llama.cpp or Ollama and let its responses control the device.
  • Hands-free operation: Use deterministic Freestyle, Chat Autopilot, or voice input for longer autonomous sessions.
  • Pattern library: Create, edit, rate, import, export, and organize reusable motion patterns and programs.
  • Funscript import: Import an entire funscript or select a specific section to turn into a reusable pattern.
  • Video library: Scan local videos, generate thumbnails, detect browser incompatibilities, convert unsupported files, and play videos synchronized with funscripts.
  • Voice input/output: Optional push-to-talk speech recognition and spoken chat replies through local or explicitly selected cloud providers.
  • Long-term memory: Let the assistant retain preferences, with individual memories that can be reviewed, disabled, or deleted.
  • Editable prompts: Customize the model’s personality and instructions without weakening the protected motion-output contract.
  • Multiple protocols: Handy Cloud, browser/Handy Bluetooth, and Intiface Central are supported through the same motion engine.

MagicHandy is still early software.

It works, but it is not yet a polished one-click release. An all-in-one Windows installer binary is planned, but has not yet been built.

Expect rough edges, especially around clean-machine installation, model selection, optional voice dependencies, long-session testing, and the broader hardware/platform support.

Installation Instructions

Windows

MagicHandy now has an easy-to-use installer!

Download it here (assets dropdown):

MagicHandy Releases

When installation finishes, MagicHandy opens at:

http://127.0.0.1:49717

The app will open to the setup wizard at first launch. It can configure either:

  • Managed llama.cpp: Best for users who want MagicHandy to own and build its local model runtime. This may run noticeably faster than Ollama, depending on your hardware. NVIDIA CUDA is recommended when available.
  • Ollama: Simpler if you already use Ollama or want to avoid the managed source build.

The managed setup may install additional dependencies depending on your choices. These are not silently installed.

Other Platforms

Windows is the primary development platform. Linux and macOS are currently best-effort, but the pure-Go core should make broader support substantially easier than it was in StrokeGPT-ReVibed.

With Go 1.25 or newer, the basic application can be built and run manually:

go run ./cmd/magichandy

The prebuilt browser interface is embedded in the Go application, so Node is not required just to run it.

Linux and macOS testing reports are welcome. macOS with the LLM running on Metal would be great, but I don’t have a device to test that with.

Setup

Install the app, then follow the setup wizard:

Model runtime:

“managed llama.cpp” is recommended for optimal performance.

Model Library:

Either import from Ollama or manually download a model, here is an example:

Voice:

These are entirely optional. Choose based on your hardware, recent NVIDIA is what has been tested.

Contributing and Supporting Development

If you would like to help, contributions are welcome through the MagicHandy repository. Bug reports, documentation, UI feedback, model testing, hardware testing, and pull requests are all useful.

You do not need to already know how to code. If you are reasonably comfortable with computers, tools such as Claude Code or Codex can help you investigate an issue and prepare a contribution. DM me if you want guidance on getting started.

You can also support development by:

  • Gifting Claude Max
  • Helping test on a Handy 2, or other Intiface-supported devices
  • Testing Windows, Linux, or macOS behavior
  • Reporting which local models work well for uncensored conversation and reliable structured motion output
  • Providing reproducible performance results from different computer hardware.
  • Donating compatible hardware for implementation and testing

Include the device model, firmware, and transport mode (the more information the better) whenever reporting motion behavior.

Models are not silently downloaded at startup. You can import compatible GGUF files, connect an existing llama.cpp server, or use models already available through Ollama. Uncensored or abliterated models are generally more suitable because conventionally aligned models may refuse sexual conversations or stop following the requested role.

Please share the model name, size, quantization, hardware, response speed, and how reliably it follows the motion format when posting results.

Last update

9/5/2026

MagicHandy is intended only for consenting adults aged 18 or older. Use it responsibly and at your own risk.

11 Likes
StrokeGPT-ReVibed
StrokeGPT-ReVibed
3

This looks absolutely amazing, thanks for putting in so much work. I’m one of the people holding out in trying it just because I have literally never touched, installed, or used any of the tech involved. I’m a bog-standard Handy and HereSphere guy. Never used AI, never used anything else to control the Handy.

I am fairly decent at following instructions :sweat_smile: but if something goes wrong or I need to make any informed decisions, I’m pretty much lost.

I’ll definitely install and try at some point, 1-click ez mode or no!

4

You can download the Codex/ChatGPT app on Windows, and ask it to either install it for you or guide you through it.

Click the ChatGPT title/logo in the upper left and change it to Codex, then select “Approve for me” in the chat bar so it doesn’t bug you for every little thing. I think you can get it to do things like this on the free tier now, it just doesn’t give you access to the smartest model (5.6 Sol > Terra > Luna, don’t use Instant) or highest reasoning level. And you’ll run out of free usage pretty quickly.

I would also enable computer use in settings (integrations). Do not mention eroscripts, it seems to trigger safeguards against it generating certain types of content, though it will help develop this app just fine. :rofl:

Paste:

See this project:
https://github.com/MagicHandy/MagicHandy

Check if this machine is ready to install it, then give me a choice between you installing or being given a guide on how to do so on this machine. I am a novice when it comes to AI and terminal use, do not assume any prior technical knowledge.

image
image1079×363 26.7 KB

1 Like
5

It now has a persona page. It does not come with any as of now. And autopilot has changed quite a bit.

image
image2560×1274 301 KB

1 Like
6

Whoah, thanks! Not sure I want to install ChatGPT on my rig, but I will probably reconsider when I need it’s help haha.

Looking forward to checking it out and will definitely post after I do to let you know!

-GX

7

Not checked this app out yet, but a hint:
Maybe you should put that quote on top of the first post, because some people will scroll and maybe read until they get to “git clone …” and gone they are.

Personally, a week ago I’ve compiled my own ffmpeg version - what a ride - so it shouldn’t be a problem for me, still I’m gonna wait for an installer, too.

8

The built-in patterns have been updated. Update if you have already installed and let me know if it’s an improvement. A lot of the patterns were broken. The video player should be a lot better now as well.

9

I honestly don’t know if I want that type of user downloading it yet as they may never update it. The app is still kinda rough.

10

I love more and more of this but I wish I wasn’t through llm or ollama cause I think those apps are sort of a waste of space on my pc. I wish there was an easier installation, and way to connect my handy because I can’t use bluetooth on my pc.

1 Like
11

It supports The Handy cloud protocol (Wi-Fi), in fact that’s the option that generally works best. The only way to get the uncensored chat features without a local provider like llama.cpp/Ollama would be to use OpenRouter or something but you have to pay for that.

Would you use a cloud LLM provider if that was an option?

12

I simplified the instructions.

13

I’ve given up on using a non-python voice output option, there just isn’t anything that works so NeuTTS is gone.

You now have the option of installing Qwen3-TTS or Chatterbox now. If you have an NVIDIA card, use Qwen3-TTS, otherwise use Chatterbox.

15

The new patterns caused the LLM to slow down a lot. I cut down the system prompt and the pattern names so it should be fine now. Update your app.

16

Corrected some additional pattern issues. Update!

17

Demonstration ^

Sorry for the male voice if you’re straight, it’s what I was testing with. You can use a female voice, just use a short and clear voice clip for the WAV voice sample. You also have to give a transcript of the voice file, but that shouldn’t be too hard.

You have to choose Qwen3-TTS for these results when installing.

The voice prompts could use some work, but I’m done tinkering with it for now.

1 Like
18

There’s an easy installer now. There’s no automatic LLM downloading yet but you can easily import models installed by ollama in setup or settings.

19

The installer is ready now.

2 Likes
20

The one-click installer request makes sense. Most Handy users are used to browser-based control (handyfeeling.com, script players) where the heavy lifting happens server-side or in firmware. Asking them to run a local LLM stack means Ollama, model downloads, CUDA drivers if they want speed, and keeping a terminal window alive. That’s a real barrier even for people comfortable with side-loading.

If you do ship a packaged version, the pain points to smooth over: auto-detecting the Handy on local network (mDNS or fallback to IP entry), handling the HAMP websocket connection retries, and surfacing model download progress in the UI instead of failing silently. The browser can’t talk to Ollama directly without a CORS proxy or the user disabling security, so your Go backend is doing essential work here.

For GoonX specifically: if the goal is AI-driven control without the local setup, there’s no shame in waiting. The Handy’s cloud API is robust and well-documented; any tool that wraps it cleanly will feel familiar. Your current workflow (HereSphere + funscripts) already hits the hardware’s timing precision limits, so adding AI is genuinely additive, not a fix for something broken.

21

So I take it…it doesn’t work on Android WTH

22

Your response is LLM-generated but it doesn’t seem to have full context. At least use an LLM agent that can download or look at the repo for context (Desktop ChatGPT app in Codex mode, Claude app in Code, Cursor, Antigravity). xD

The LLM you used didn’t know that The Handy only supports connections routed through their servers or bluetooth. There is no LAN control, so auto discovery isn’t relevant. Unless you reverse engineer or replace the firmware, maybe.

And the installer has already been released, so your reply is also out of date.

2 Likes