HKUDS · 2026-04-08 · notable
OpenHarness — Open-Source Agent Infrastructure with Built-in Personal Agent
Lightweight open-source agent harness providing tool-use, skills, memory, and multi-agent coordination. Includes Ohmo, a personal AI agent with 43 tools and persistent cross-session memory. 8.9k stars.
A lightweight, MIT-licensed agent harness that gives researchers and builders the core infrastructure for tool-use, memory, and multi-agent coordination.
Key specs
| License | MIT |
|---|---|
| GitHub stars | 8.9k |
| Built in tools | 43 |
What is it?
OpenHarness is an open-source Python framework from the Hong Kong University Data Intelligence Lab (HKUDS) that provides foundational infrastructure for building AI agents. It ships with 43 built-in tools, persistent memory, context compression, and multi-agent coordination. It also includes Ohmo, a personal agent app built on top of the harness with its own workspace and gateway.
How does it work?
The core is an agent loop engine that cycles through streaming tool calls. It supports multiple LLM providers (Claude, OpenAI, Gemini, Kimi), integrates with MCP servers via HTTP transport, and handles context compression automatically. v0.1.6 added auto-compaction that preserves task state across context window resets. A React-based terminal UI provides the developer interface.
Why does it matter?
Understanding how production AI agents work under the hood is difficult when the only options are closed-source products. OpenHarness gives researchers and builders a transparent, extensible codebase to experiment with agent architectures, benchmark tool-use patterns, and prototype custom agents without building from scratch.
Who is it for?
AI researchers, agent framework developers, and builders who want a transparent agent codebase to extend.