Technology RadarTechnology Radar

OpenHarness

agentopen-sourcemulti-agentframework
Assess

OpenHarness is an open-source Python agent harness from the Hong Kong University Data Science Lab (HKUDS) — a transparent, inspectable implementation of core agent infrastructure (tool-use, skills, memory, multi-agent coordination, permissions) that can wrap any LLM provider. It ships with ohmo, a personal assistant that connects to Feishu, Slack, Telegram, and Discord on top of the same harness.

What It Is

OpenHarness provides the scaffolding layer that turns a language model into a functional agent: a streaming tool-call loop, a toolkit of 43+ tools (file I/O, shell, web search, MCP), a skills system that loads domain knowledge from Markdown files, persistent memory (MEMORY.md), multi-level permission governance, and a React/Ink terminal UI. The codebase is organized into 10 explicit subsystems with clean boundaries, making it well-suited for researchers who want to inspect or modify how a production-grade harness works.

The project is explicitly positioned as a research and community tool — an open reference implementation of the same architectural patterns used by Claude Code, Codex CLI, and Goose.

Architecture

Subsystem Role
Engine Streaming agent loop: query → stream → tool-call → execute → loop
Tools 43+ registered utilities (file, shell, search, web, MCP, task)
Skills On-demand Markdown knowledge files, matching by topic
Permissions Multi-level modes (Default / Auto / Plan), path rules, command denials
Hooks PreToolUse / PostToolUse lifecycle events
Memory Persistent MEMORY.md with auto-compaction for long sessions
MCP Model Context Protocol client for external tool servers
Coordinator Subagent spawning, team registry, background task lifecycles
Commands 54 built-in workflow commands
TUI React/Ink terminal interface

Multi-Provider Support

OpenHarness is designed for provider portability — a differentiator versus Claude Code (Anthropic-native) or Codex CLI (OpenAI-native):

Profile type Supported providers
Anthropic-compatible Claude (official), Moonshot/Kimi, Zhipu GLM, MiniMax
OpenAI-compatible OpenAI, OpenRouter, DeepSeek, Groq, Ollama, GitHub Models
Subscription bridges Claude Code CLI, Codex CLI, GitHub Copilot (OAuth device flow)

The subscription bridge capability is notable: teams with Claude Code or Codex subscriptions can run OpenHarness on top of them without separate API keys.

Compared to Claude Code and OpenCode

OpenHarness Claude Code OpenCode
License MIT Proprietary MIT
LLM support Multi-provider Claude (primary) 75+ providers
Origin Academic (HKUDS) Anthropic Community
Focus Research / inspection Production productivity Developer productivity
Sandbox Host environment Host environment Host environment
MCP support Client (built-in) Client (built-in) Supported
Personal assistant ohmo (Slack/Feishu/Discord) None None

Why Assess

OpenHarness earned 9,300+ GitHub stars and 1,600+ forks within two weeks of its initial release (v0.1.0, April 1, 2026), signalling strong community interest. The architecture is well-documented, the test suite covers 114 unit/integration tests and 6 CLI E2E suites, and plugin compatibility with the anthropics/skills and claude-code/plugins ecosystems has been verified.

The reasons not to Trial yet:

  • Too new: Two weeks from initial release (v0.1.6 as of April 10, 2026). Production stability is unproven at scale.
  • Research framing: Explicitly positioned for researchers and community builders, not production engineering teams.
  • No named enterprise adopters: The showcase and documentation do not cite any production deployments.
  • Academic origins: HKUDS is a university lab; long-term maintenance trajectory is less certain than corporate-backed alternatives.

Assess means: understand how it works, experiment with the provider-portability and subscription-bridge features, and track whether community adoption converts into production use cases over the next 2–3 months.

Getting Started

# Linux/macOS/WSL
curl -fsSL https://raw.githubusercontent.com/HKUDS/OpenHarness/main/scripts/install.sh | bash

# Configure
oh setup

# Run interactively
oh

# Single prompt (headless)
oh -p "Review this repo and identify the highest-risk bug"

# JSON output for pipelines
oh -p "List all TODO comments" --output-format json

Key Characteristics

Property Value
Interface React/Ink TUI + CLI
License MIT
LLM backends Anthropic-compatible + OpenAI-compatible
Initial release April 1, 2026 (v0.1.0)
Latest release v0.1.6 (April 10, 2026)
GitHub stars ~9,300 (April 2026)
Tests 114 unit/integration + 6 CLI E2E suites
Origin HKUDS (HKU Data Intelligence Lab)
GitHub HKUDS/OpenHarness

Sources