Avdpro/ai2apps

What it solves

AI2Apps provides a local-first platform for deploying and managing AI applications, agents, and services on personal AI nodes. It bridges the gap between raw model runtimes and usable software by adding a management layer for user identity, durable agent execution, and secure tool access, allowing users to run complex AI workflows locally while maintaining data privacy.

How it works

The platform operates through a layered architecture:

• Apps: Handle the user interface and persistent state (e.g., Chat, Coder).
• Agents: Managed by a server-side "Harness" that ensures execution is durable, restart-aware, and capable of handling tool calls and user approvals.
• Services: Provide stable, versioned identities for capabilities like browser control, document parsing, and model access, removing the need for fixed URLs.
• Runtime: A replaceable backend (initially using oMLX for Apple Silicon) that handles the actual model inference.

It also includes a specialized "Flesh engine" for large Mixture-of-Experts (MoE) models, optimizing memory and I/O through expert-major SSD storage and adaptive caching.

Who it’s for

• Developers building local AI applications and agents.
• Users with Apple Silicon Macs wanting to run large MoE models (like DeepSeek V4) with optimized local inference.
• Teams needing a secure, multi-user environment for local AI nodes with role-based access and package verification.

Highlights

• Durable Agent Execution: Supports pausing, resuming, and recovering agent runs with replayable events.
• Local-First Infrastructure: Keeps session data local while allowing remote access and cloud model bridging.
• Secure Package System: Uses signed and content-addressed packages for apps, agents, and services with built-in rollback and Safe Mode.
• MoE Optimization: Implements advanced routing and caching strategies to handle massive models on limited local hardware.