Pneuma: An AI-Generated Desktop Environment Where Software Materializes from Descriptions
Pneuma is a desktop computing environment where software doesn't exist before you need it. You boot to a blank screen with a prompt, describe what you want—like a CPU monitor, game, notes app, or data visualizer—and a working program materializes in seconds.
How It Works
Your input goes to an LLM, which generates a self-contained Rust module. That gets compiled to WebAssembly in under a second, then JIT-compiled and executed in a sandboxed Wasmtime instance. Everything is GPU-rendered via wgpu (supporting Vulkan, Metal, and DX12). If compilation fails, the error is automatically fed back for correction, with a ~90% first-attempt success rate.
Architecture and Features
- Microkernel architecture: Agents run in isolated WASM sandboxes with a typed ABI for drawing, input, storage, and networking
- Agent persistence: Once generated, agents persist—you can reuse them, they communicate through IPC, and you can share them through a community agent store
- System stability: An agent crash can't bring down the system
- Agent capabilities: Agents can run side by side, persist to a local store, and be shared or downloaded from the community store
Current Status and Future Goals
Currently runs as a desktop app on Linux, macOS, and Windows. The longer-term goal is to run on bare metal and support existing ARM64 binaries alongside generated agents, creating a full computing environment where AI-generated software and traditional applications coexist. Built entirely in Rust.
Philosophy
The creator built this because they believe the traditional software model—find an app, install it, learn it, configure it—creates unnecessary friction. If a computer can generate exactly the tool you need when you need it, and keep it around when it's useful, why maintain a library of pre-built software at all?
A free tier is available (no credit card required), and there's a video on the landing page showing it in action. The creator is interested in feedback on the concept, UX, and whether this is something developers would actually use.
📖 Read the full source: HN LLM Tools
👀 See Also
Single-page chatbot interface for locally running Gemma 4 26B A4B
A developer built a single HTML page chatbot that connects to Gemma 4 26B A4B running locally with 32K context window at 50-65 tokens/second, sharded between a 7900 XT and 3060 Ti GPU. The interface includes full streaming, Markdown rendering, and parameter controls.
Claude Dispatch Beta: Setup Tips and Initial Impressions
A developer shares their experience setting up Claude's Dispatch beta on a Mac Mini, highlighting the need for constant uptime, specific success criteria, and aggressive permissions with Computer Use enabled.
Terrarium: Open-Source Sandbox for Agentic Environments with Time Machine Rewind
A versatile sandboxing solution for running multiple AI agents securely on any VPS or cloud. Features isolated worlds, reverse-proxy management, GUIs, and a time machine to rewind container state.
Applying Claude Code's Architecture to Local 9B Models: Key Findings and Optimizations
A developer extracted architectural patterns from Claude Code's leaked source code and applied 10 optimizations to qwen3.5:9b running locally on an RTX 5070 Ti. The key discovery was that qwen3.5:9b has native structured tool_calls, and the biggest limitation for 9B models is self-discipline in knowing when to stop exploring and start producing output.