eastreams/loong

What it solves

Loong provides a secure, extensible, and stable foundation for building "vertical AI agents"—specialized agents designed for real-world scenarios. It addresses the need for a controllable agent base that can handle long-horizon workflows, execute complex compound tasks, and implement closed-loop improvements without becoming a tangled mess of configurations.

How it works

Built in Rust, Loong uses a modular architecture that separates the core kernel from extensions. It employs a layered execution model and a strict acyclic dependency graph across its crates to ensure stability. The system is organized into distinct components:

• Providers: Connects to various LLM backends (e.g., OpenAI, Volcengine).
• Channels: Manages communication surfaces (e.g., Lark/Feishu).
• Gateway & Runtime: Operates within explicit boundaries to manage tools, memory, approvals, and policies.
• Kernel: A governed core that separates stable contracts from the runtime substrate.

Who it’s for

Loong is designed for both beginners who want a quick start via a guided onboarding process and power users/developers who need a professional-grade, sustainable base to build and extend complex AI agents.

Highlights

• Extensive Out-of-the-Box Support: Includes over 42 built-in providers and 25 channels.
• High Controllability: Uses named operator shells to group functions (sessions, skills, channels, etc.) rather than a flat command list.
• Secure Base: Implements explicit runtime boundaries for provider selection, memory, and audit logs.
• Modular Design: Providers, tools, and policies live outside the kernel, allowing them to be compiled and composed as needed.
• Developer Friendly: Features a TUI (Terminal User Interface) and a guided onboard command to avoid manual TOML editing.