DemiArch

Overview

DemiArch is a local-first, open-source AI application builder written in Rust. It generates production-ready code through conversational AI interfaces, transforming project ideas into complete codebases while keeping all data local and under user control.

Status: Early Development | Open Source (AGPL-3.0)

Core Philosophy

DemiArch is built on four foundational principles:

Code Generation Focus - Generates code for local repositories; users run their own projects
Local-First Architecture - All data stays local using SQLite; no accounts or telemetry required
Explicit Operations - No automatic git operations or background processes without user consent
User Ownership - Users can freely edit generated code; AI respects modifications during regeneration

Problem Solved

DemiArch addresses key limitations of cloud-based AI coding tools:

Privacy - No data leaves your machine; no accounts required
Control - You own the code and can edit it freely
Cost Transparency - Direct LLM API usage with visible per-model pricing
No Lock-in - Open source with standard code output

Key Features

Conversational Discovery

Natural language project description
Auto-generated PRD, architecture docs, and UX specifications
Phase planning with user stories
Iterative refinement through conversation

Multi-Framework Support

Frontend: Next.js, React, Vue, Flutter
Mobile: iOS, Android
Backend: FastAPI, Go
Extensible framework definitions

Russian Doll Agent System

Hierarchical 3-level agent delegation:

Orchestrator - High-level planning and coordination
Specialists - Domain-specific implementation
Workers - Atomic code generation tasks

Learned Skills System

Auto-extracts debugging knowledge from sessions
Persists solutions for future similar problems
Progressive disclosure for token-efficient context retrieval

Smart Model Routing

Dynamic model selection using reinforcement learning
Per-model pricing awareness
Daily budget controls
Cost optimization without sacrificing quality

Multi-Project Support

Work on multiple projects concurrently
Cross-project context awareness
Isolated project databases

Technical Architecture

Core Stack

Language: Rust for performance and safety
Database: SQLite with vector extensions for local storage
LLM Backend: OpenRouter API for model flexibility

Interfaces

CLI: Fast command-line interface for scripting
TUI: Rich terminal UI built with Ratatui
GUI: Planned Tauri + React + TypeScript desktop app

Project Structure

Modular Rust crate architecture:

demiarch-core - Core library with all business logic
demiarch-cli - Command-line interface
demiarch-tui - Terminal user interface
Plugin system with WASM sandboxing for extensibility

Local-First Data

SQLite database per project
Vector extensions for semantic search
No cloud sync required
Full data portability

Architecture Highlights

Token Efficiency

Progressive disclosure system retrieves only relevant context, minimizing token usage and costs while maintaining code quality.

Plugin System

WASM-sandboxed plugins allow safe extensibility without compromising system security.

Cost Management

Real-time cost tracking per model
Daily budget limits
Model routing optimized for cost/quality balance
Full visibility into API spending

Technical Excellence

DemiArch demonstrates mastery of:

Systems Programming: Rust, memory safety, performance optimization
Local-First Architecture: SQLite, offline-capable, data ownership
AI Engineering: LLM orchestration, prompt engineering, agent hierarchies
TUI Development: Ratatui, terminal interfaces, keyboard-driven UX
Plugin Architecture: WASM sandboxing, extensibility, security
Open Source: AGPL-3.0, community-driven development

Status: Early Development License: AGPL-3.0 Repository: GitHub