mkbhardwas12/prismos-ai

What it solves

PrismOS-AI provides a completely private, offline AI environment on a local machine. It eliminates the need to send sensitive data to remote servers by running LLMs locally, ensuring that no bytes leave the device during standard operation.

How it works

The application uses a Tauri-based desktop interface (React/Rust) that connects to a local Ollama daemon for model inference. It implements a local RAG (Retrieval-Augmented Generation) system using TF-IDF for document extraction and a local SQLite knowledge graph to persist information across sessions. For security, it runs AI agents within a wasmtime sandbox with strict memory and CPU limits, and enforces a strict Content Security Policy (CSP) to block unauthorized network requests.

Who it’s for

Users who require high privacy for their AI interactions, those working with sensitive documents (PDF, DOCX, PPTX, XLSX), and developers who want a local-first AI assistant with agentic capabilities.

Highlights

• Zero-Network Default: Enforces offline operation via OS-level CSP, ensuring data stays on the laptop.
• Local Knowledge Graph: Stores learned concepts and answers in a local SQLite database for long-term memory.
• Sandboxed Agents: Features 8 specialized agents (e.g., reasoner, memory keeper, finance) running in a WASM container for safety.
• Dynamic Model Routing: Automatically switches between text and vision models based on the input type (e.g., attaching an image).
• Hardware-Aware: Provides model recommendations based on the user's local hardware specifications.