Anything AI

A Mixture of Experts AI Assistant for your Desktop Terminal

Built by BlueFireStudios

---

What is Anything AI?

Anything AI is a free, open-source terminal AI assistant that routes your prompts to the best specialized AI model for the job. Instead of using one model for everything, it uses a Mixture of Experts architecture, a fast router model reads your intent and dispatches it to the most capable specialist model available.

No subscriptions. No API keys to buy. Runs on Hugging Face's free inference tier.

---

Features

• Intelligent Intent Routing : Deepseek R1 Distilled 70B classifies every prompt and sends it to the right expert
• Specialist Models for each task:
  • code_create / code_edit → Qwen3 Coder 480B
  • reasoning / math → DeepSeek R1 Distill 70B
  • conversation / roleplay → Qwen2.5 72B
  • agentic_task → Qwen2.5 Coder 32B
  • file_generate → Qwen3 Coder 480B
• Agentic Tasks : open and close apps, launch websites directly from chat
• File Generation : create PowerPoint, Word, PDF, Excel, CSV, and text files from a single prompt
• Image Input : attach images via Insert key or Explorer file picker, analyzed before being sent to the model
• Smart Memory : keeps the last 5 exchanges raw, compresses older context automatically in the background
• Multi-Token Fallback : chains up to 3 Hugging Face tokens so you never hit a rate limit mid-session
• Code Saving : generated code is saved directly to your output folder with AI-suggested filenames
• TUI Interface : clean terminal UI with typewriter output, colorama coloring, and ASCII header

---

Requirements

• Python 3.10+
• Windows (agentic features use Windows-specific app control)

Dependencies

pip install huggingface_hub
pip install AppOpener
pip install python-pptx
pip install python-docx
pip install reportlab
pip install openpyxl
pip install requests
pip install Pillow
pip install pynput
pip install colorama

---

Setup

1. Download the .exe file [EXPERIMENTAL]

Just go to the releases tab! Though be cautious as it is not ready for shipping yet.

2. Set your Hugging Face token(s) as environment variables

You need at least one free Hugging Face account. Create a token at https://huggingface.co/settings/tokens

On Windows:

setx HF_ACCESS_TOKEN "hf_yourtoken"
setx HF_ACCESS_TOKEN_2 "hf_yourtoken2"
setx HF_ACCESS_TOKEN_3 "hf_yourtoken3"

Tokens 2 and 3 are optional but recommended, the app automatically falls back to the next token if one hits its rate limit.

3. Run

Open the Anything AI.exe file!

---

Usage

Just type naturally. Anything AI figures out what you need.

What you type	What happens
write me a snake game in python	Qwen3 Coder generates the code, offers to save it
fix the bug in the last code	Qwen3 Coder edits with full context of previous code
what is the derivative of x squared	DeepSeek R1 solves it step by step
open spotify	AppOpener launches Spotify
go to github.com	Opens in your default browser
make a presentation about black holes	Generates a .pptx with slides and Wikimedia images
what's in this image	Press Insert to attach, model describes it

Keyboard Shortcuts

Key	Action
Insert	Open file explorer to attach any supported file (image files supported by PIL, Certain text and Microsoft Office files, and code files like .py, .java, .cpp etc.
Enter	Send prompt (Straifghtforward, isn't it?)

Output Files

All generated files and saved code are stored at:

C:\Users\[YourUsername]\Anything AI\output\

And will be automatically opened on your PC

Architecture

User Prompt
     │
     ▼
DeepSeek R1 70B (Router)
     │
     ├── code_create  ──► Qwen3 Coder 480B
     ├── code_edit    ──► Qwen3 Coder 480B
     ├── reasoning    ──► DeepSeek R1 70B
     ├── math         ──► DeepSeek R1 70B
     ├── conversation ──► Qwen2.5 72B
     ├── roleplay     ──► Qwen2.5 72B
     ├── agentic_task ──► Qwen2.5 Coder 32B
     └── file_generate──► Qwen3 Coder 480B

Memory System

Context is stored as a rolling list. The last 5 exchanges are kept raw for immediate reference. Anything older is compressed by Llama 3.1 8B in a background thread so memory never bloats without losing important details.

Long term context is saved periodically, in Anything AI\memory\memory.txt

Token Fallback Chain

Every API call tries client1 → client2 → client3 in sequence. If one token is rate limited or expired, the next one takes over transparently.

---

Limitations

• Max code output is ~2000 tokens per generation, but it depends on your Hugging Face account tier.
• Is currently Windows-only

---

License

MIT License: free to use, modify, and distribute.

---

Made by BlueFireStudios

This README was generated by Anything AI, and edited by me.