Experimental AI NPC desktop built with Tauri + SvelteKit. Local-first, self-hosted gateways, and agent gameplay experiments.
Most AI tools today feel like browser chat wrappers.
I wanted to explore a different direction — something closer to a self-hosted AI desktop environment where agents, documents, and even games exist inside one local workspace.
That experiment became openClaw Desktop.
👉 https://github.com/stormixus/openClaw-Desktop
🧠 The Idea
Instead of designing another chat UI, I started asking:
- What if AI behaved more like an NPC inside a desktop?
- What if local gateways handled models and sessions instead of cloud-only APIs?
- What if gameplay, documents, and agents shared the same runtime?
The result is an experimental AI NPC Desktop Platform built with:
- Tauri 2 (Rust backend)
- SvelteKit frontend
- Local-first gateway architecture
This isn’t a polished product — it’s a playground for exploring AI UX.
🎭 NPC-Style Interaction
One of the core experiments is NPC mode.
Messages can include simple directives like:
[face:happy]
[act:thinking]
Instead of static chat bubbles, the AI feels more like a character reacting inside the desktop.
The goal isn’t realism — it’s exploring how interaction changes when AI feels present rather than embedded in a webpage.
🔌 Self-Hosted Gateway Architecture
Rather than locking into a single provider, openClaw connects to self-hosted gateways.
This allows:
- Local LLM setups
- Remote model switching
- Multiple active sessions
- Agent vs Agent experiments
It’s basically a thin desktop shell around flexible model backends.
📄 Forge — Agent-Driven Document Workspace
Another experiment is a workspace called Forge.
Instead of editing documents manually, the flow looks more like:
- Open file
- Ask agent for changes
- Review staged patch
- Apply or discard
Current experiments include:
- OCR extraction
- Markdown / JSON editing
- Spreadsheet handling
- AI-assisted modifications
The focus is on workflow, not just generation.
🎮 Why Are There Games?
This part surprises people 😄
The built-in game hub started as a way to test agent reasoning loops.
Games like:
- Chess
- Go
- Poker
- Puzzle boards
support:
- Player vs Agent
- Agent vs Agent
It turns out games are a really good testbed for interaction design and UI experiments.
🖥️ Why Tauri?
I chose Tauri because:
- Native desktop feel without Electron weight
- Easy Rust ↔ frontend bridge
- Good fit for local-first architecture
The desktop shell stays lightweight while still supporting complex UI experiments.
⚠️ Current Status
This project is still experimental:
- Architecture is evolving fast
- Plugin runtime is in progress
- NPC interaction is more of a UX prototype
I’m sharing it early because I’m curious how others think about:
- Local-first AI interfaces
- NPC-style interaction
- Self-hosted gateway designs
🚀 Repo & Demo
GitHub:
https://github.com/stormixus/openClaw-Desktop
GIF previews and demos are in the README.
If you’re experimenting with local LLM setups, Tauri apps, or alternative AI UI ideas — I’d love to hear your thoughts.
Top comments (0)