Summary

LoongClaw is a Rust-based AI agent infrastructure designed as a foundation for vertical, team-facing agents. It prioritizes security, extensibility, and evolvability, aiming to move beyond general assistants to support collaborative workflows and potentially embodied intelligence. The project benefits teams by providing a stable, governed core for building specialized AI solutions with clear extension points.

How It Works

Built in Rust, LoongClaw employs a modular architecture with distinct "execution planes" (connector, runtime, tool, memory) and a separate "control plane" (ACP). This design facilitates vertical specialization by allowing planes to be replaced rather than modifying the core kernel. Governance primitives—capability tokens, policy decisions, approvals, and audit events—are integrated into critical execution paths, fostering a team-centric system. Its context engine and tool surface are designed for flexibility and runtime accuracy, enabling specialized agents to evolve on a stable, extensible Rust base.

Quick Start & Requirements

• Install: Use the provided shell script: curl -fsSL https://raw.githubusercontent.com/loongclaw-ai/loongclaw/main/scripts/install.sh | bash -s -- --onboard (Linux/macOS) or the PowerShell equivalent.
• Build from Source: cargo install --path crates/daemon or via install scripts.
• Prerequisites: Rust toolchain for source builds; API keys for AI providers.
• Links: Installation scripts are available at https://raw.githubusercontent.com/loongclaw-ai/loongclaw/main/scripts/install.sh and https://raw.githubusercontent.com/loongclaw-ai/loongclaw/main/scripts/install.ps1.
• First Steps: Run loongclaw onboard, set provider credentials (e.g., export PROVIDER_API_KEY=...), and interact via loongclaw ask --message "...".

Highlighted Details

• Governance-native execution: Capability tokens, policy decisions, approvals, and audit events are core to execution paths.
• Explicit execution planes: Connector, runtime, tool, and memory planes offer clear core/extension registration for specialization.
• Separate control plane: ACP manages backend, binding, runtime, and store modules for enhanced routing and coordination.
• Shapeable context: Context engine supports hooks for lifecycle management (bootstrap, ingest, after_turn, etc.).
• Runtime-truthful tool surface: Tool catalog details risk classes and approval modes for accurate capability representation.
• Migration-aware setup: onboard detects existing setups; loongclaw migrate CLI facilitates discovery, planning, application, and rollback.
• Multi-surface delivery: Supports CLI, Telegram, Feishu/Lark, and Matrix channels.

Maintenance & Community

The project encourages contributions via a dedicated guide and lists areas where help is especially welcome. However, specific links to community channels (e.g., Discord, Slack), active maintainer details, or sponsorship information beyond a placeholder are not detailed in the provided README.

Licensing & Compatibility

LoongClaw is licensed under the MIT License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Some ecosystem components are described as "architecture direction" rather than fully realized product features. The long-term vision includes hardware, robotics, and embodied intelligence, which may represent future development rather than current capabilities.