HiClaw provides an open-source system for deploying and managing AI agent teams, enabling multi-agent collaboration with human oversight via instant messaging. It targets users needing to orchestrate AI agents efficiently and securely, offering a "chief of staff" manager for task delegation and progress monitoring.

How It Works

A central Manager Agent, acting as an AI chief of staff, orchestrates Worker Agents. Tasks are assigned conversationally, and all interactions occur within Matrix rooms, providing full transparency and allowing human intervention at any stage. HiClaw employs a robust security model where Workers only handle consumer tokens, while the Higress AI Gateway securely manages sensitive API keys and GitHub Personal Access Tokens, preventing credential leakage.

Quick Start & Requirements

Installation is streamlined via a single curl | bash script for Linux/macOS or a PowerShell script for Windows.

• Prerequisites: Docker Desktop (Windows/macOS) or Docker Engine (Linux)/Podman Desktop.
• Resource Requirements: Minimum 2 CPU cores and 4 GB RAM; 4 CPU cores and 8 GB RAM recommended for multiple Workers due to OpenClaw's memory usage.
• Documentation: Quick Start (docs/quickstart.md), Architecture (docs/architecture.md), Manager Guide (docs/manager-guide.md), Worker Guide (docs/worker-guide.md), Development (docs/development.md).

Highlighted Details

• Security by Design: Workers operate with limited consumer tokens; real credentials are managed centrally by the Higress AI Gateway.
• IM-Based Collaboration: Utilizes Matrix for real-time, transparent communication and human-in-the-loop control within familiar chat interfaces.
• Skills Ecosystem: Workers can dynamically access over 80,000 community skills from skills.sh.
• OpenClaw Differentiation: Offers distributed container deployment, conversational agent creation, enhanced security, built-in visibility, and mobile access compared to native OpenClaw.

Maintenance & Community

Officially open-sourced on March 4, 2026. Community support is available via Discord, DingTalk Group, and WeChat Group. The roadmap includes developing lightweight Worker runtimes (CoPaw, ZeroClaw, NanoClaw) to reduce memory footprint, a Team Management Center for enhanced observability, and universal MCP service support for broader tool integration.

Licensing & Compatibility

Licensed under the Apache License 2.0, which generally permits commercial use and integration into proprietary systems.

Limitations & Caveats

Current Worker implementations rely on OpenClaw, which exhibits relatively high memory consumption (~500MB per Worker). Future roadmap items aim to address this by introducing alternative, more lightweight runtimes. As a recently open-sourced project, it may be subject to rapid development and potential API changes.