OrionBot

Get up and running with Claude, GPT, Gemini, Llama, and other models through a native Windows desktop app —

Note: This is a demo / preview build, not the final product. Things will change. NVIDIA NIM and OpenRouter have been tested; the local-model (Ollama / LM Studio) path has not been tested yet — use it at your own discretion for now.

---

Download

Grab OrionBot.exe from Releases and run it. No Python, no dependencies, nothing else to install.

---

Quickstart

1. Run OrionBot.exe
2. On first launch, pick a connection method:
   • API Key — paste your key, the provider is auto-detected (sk-ant- → Anthropic, sk-or- → OpenRouter, nvapi- → NVIDIA)
   • Local model — point it at Ollama, LM Studio, or your own OpenAI-compatible endpoint
3. Start chatting

No config files to touch by hand.

---

Features

• Auto-detected providers — paste a key, OrionBot figures out which service it belongs to
• Local model support — Ollama, LM Studio, or any custom endpoint, no key required
• Persistent, multi-chat — every conversation is saved as a linked Markdown file (Obsidian-style), can be starred, browsed, and revisited
• Vault — a persistent, Markdown-based memory the assistant can write notes into
• Skills, Claude Skills–compatible — drop in a SKILL.md folder, a .zip, or a GitHub repo link and OrionBot picks up the new capability
• Terminal integration — the assistant can run shell commands, gated by a confirm dialog (or an "allow all" toggle if you trust it)
• Optional Home Assistant integration — control lights, switches, and thermostats from chat
• Single .exe, no Python install required

---

Skills

OrionBot's skill system speaks the same format as Anthropic's Claude Skills — a folder with a SKILL.md file:

my-skill/
├── SKILL.md          required
├── scripts/          optional
└── references/       optional

---
name: pdf-summarizer
description: "Use this when the user wants a PDF read or summarized."
---

# PDF Summarizer

...

The description field is what OrionBot feeds to the model as context, so it can decide when the skill is relevant and ask for the full file with SKILL_OKU: <skill-name> if needed.

Add a skill three ways, right from the app:

• Drag and drop a .zip, .md, or .py file
• Paste a GitHub repo link — OrionBot pulls SKILL.md or README.md automatically
• Write one by hand (see backend/skills.py if you're scripting it)

This has been tested against real skills from Anthropic's own skill collection (docx, pdf, etc.) — they load and parse without modification.

---

Connection methods

Method	Description
API Key	Paste a key, provider is auto-detected
Local model	Ollama, LM Studio, or any OpenAI-compatible endpoint — not yet tested, use with caution
Kortex	Coming soon — OrionAGI's own model. Currently a disabled placeholder in the UI

---

Building from source

The .exe on the Releases page is built from this repo with:

cd frontend
pip install pywebview requests psutil pyinstaller
python build_exe.py

This produces dist/OrionBot.exe.

If you just want to run it from source without packaging:

cd frontend
python main.py

This works fine for development, but the title bar may briefly show the default Python/Windows icon instead of OrionBot's before the app applies its own — a pywebview quirk, not a bug. The packaged .exe sets it directly.

---

Architecture

orionbot/
├── backend/
│   ├── orion_engine.py     the engine tying everything together
│   ├── ai_client.py        multi-provider AI client + key auto-detection
│   ├── vault.py             Obsidian-style persistent notes
│   ├── conversations.py     persistent, multi-chat history
│   ├── skills.py            Claude Skills–compatible skill system
│   ├── terminal.py          command runner
│   └── home_assistant.py    home automation integration (optional)
└── frontend/
    ├── index.html           the entire UI (single file)
    ├── main.py               pywebview bridge (native window, no localhost)
    └── build_exe.py          .exe packager

The frontend runs on pywebview's edgechromium backend (the WebView2 runtime already on most Windows machines) — a direct Python↔JavaScript bridge, no HTTP server, no browser tab.

---

Roadmap

• Test and stabilize the Ollama / local-model path
• Linux and macOS builds
• Kortex — a lightweight local model built for this app specifically

---

License

MIT