OpenHarness Open Agent Framework | Generated by AI

Home 2026.04

Question: What is OpenHarness, the open-source project at https://github.com/HKUDS/OpenHarness?

Answer:

OpenHarness — Open Agent Harness (oh)

OpenHarness is an open-source Python framework developed by the HKUDS (Hong Kong University Data Science) group. It delivers core lightweight agent infrastructure covering tool-use, skills, memory, and multi-agent coordination. The project was initially released on April 1, 2026 as v0.1.0, and has quickly gained traction with over 5.7k stars and 973 forks on GitHub.

What is an “Agent Harness”?

An Agent Harness is the complete infrastructure that wraps around an LLM to make it a functional agent. The model provides intelligence; the harness provides hands, eyes, memory, and safety boundaries.

In short:

Core Architecture

OpenHarness implements the core Agent Harness pattern with 10 subsystems:

Subsystem Role
engine/ Agent loop — stream, tool-call, loop
tools/ 43+ tools (file, shell, search, web, MCP)
skills/ On-demand knowledge loading via .md files
plugins/ Extensions — commands, hooks, agents, MCP servers
permissions/ Safety — multi-level modes, path rules
hooks/ Lifecycle events (PreToolUse / PostToolUse)
commands/ 54 slash-commands (/help, /commit, /plan, etc.)
memory/ Persistent cross-session knowledge via MEMORY.md
coordinator/ Multi-agent subagent spawning and team coordination
ui/ React-based Terminal UI (TUI)

Key Features

1. Agent Loop Engine

The Agent Loop supports streaming tool-call cycles, API retry with exponential backoff, parallel tool execution, and token counting & cost tracking.

2. Tools (43+)

Tools span multiple categories: File I/O (Bash, Read, Write, Edit, Glob, Grep), Search (WebFetch, WebSearch, ToolSearch), Agent tools for subagent spawning, Task management, and MCP (Model Context Protocol) integration. Every tool has Pydantic input validation, self-describing JSON Schema, permission integration, and hook support.

3. Skills System

Skills are on-demand knowledge loaded only when the model needs them. Examples include commit, review, debug, plan, test, simplify, pdf, xlsx, and 40+ more. It is also compatible with the official anthropics/skills repo — just copy .md files to ~/.openharness/skills/.

4. Permissions & Safety

Multi-level permission modes are supported: Default (ask before write/execute), Auto (allow everything, for sandboxed environments), and Plan Mode (block all writes for review-first workflows). Path-level rules and denied command lists can be configured in settings.json.

5. Multi-Agent / Swarm Coordination

OpenHarness supports subagent spawning and delegation, team registry and task management, background task lifecycle management, and ClawTeam integration (on the roadmap).

6. Plugin System

Compatible with claude-code plugins, tested with 12 official plugins. Examples include commit-commands for git workflows, security-guidance for security warnings on file edits, code-review for multi-agent PR review, and hookify for creating custom behavior hooks.

Provider Compatibility

OpenHarness supports three API formats: Anthropic (default), OpenAI-compatible (--api-format openai), and GitHub Copilot (--api-format copilot).

Supported providers via OpenAI format include DeepSeek, Alibaba DashScope, OpenAI, GitHub Models, SiliconFlow, Groq, and local Ollama. GitHub Copilot support uses OAuth device flow — no API keys needed.

Quick Start

# One-click install
curl -fsSL https://raw.githubusercontent.com/HKUDS/OpenHarness/main/scripts/install.sh | bash

# One-command demo
ANTHROPIC_API_KEY=your_key uv run oh -p "Inspect this repository and list the top 3 refactors"

# Interactive mode
oh

# Non-interactive / scripting
oh -p "List all functions in main.py" --output-format json

Who Is It For?

OpenHarness is designed for researchers, builders, and the community who want to: understand how production AI agents work under the hood, experiment with cutting-edge tools, skills, and agent coordination patterns, extend the harness with custom plugins, providers, and domain knowledge, and build specialized agents on top of proven architecture.

Tech Stack

References:

Back Donate