OpenHarness delivers lightweight agent infrastructure, addressing core needs like tool-use, skills, memory, and multi-agent coordination. It targets researchers and developers, offering an inspectable and extensible harness to understand, experiment with, and build functional AI agents efficiently.

How It Works

The project implements the "Agent Harness" pattern, abstracting LLM intelligence into a functional agent. Its core is a Python engine featuring an agent loop for streaming tool-calls, parallel execution, and cost tracking. Key advantages include a modular architecture with 10 subsystems, a focus on pure harness functionality stripped of enterprise overhead, and a significantly smaller codebase (44x lighter than Claude Code).

Quick Start & Requirements

• Install/Run: Clone the repo, uv sync --extra dev, then uv run oh.
• Prerequisites: Python 3.10+, Node.js 18+ (optional TUI), LLM API key (e.g., ANTHROPIC_API_KEY).
• Links: Quick Start, Provider Compatibility, Showcase, Contributing, Changelog.

Highlighted Details

• Tools: 43+ tools (File I/O, Search, Notebook, Agent, Task, MCP, Mode, Schedule, Meta) with Pydantic validation, JSON Schema, permission checks, and hook support.
• Skills: On-demand knowledge loading from .md files, compatible with anthropics/skills.
• Plugins: Extensible ecosystem for custom commands, hooks, and agents, compatible with claude-code/plugins.
• Governance: Multi-level permission modes (Default, Auto, Plan) with path-level rules and command restrictions.
• UI: Interactive React/Ink TUI offering command picking, permission dialogs, and session management.
• Performance: Notably lightweight, with 11,733 lines of TypeScript code compared to 512,664 for Claude Code.

Maintenance & Community

OpenHarness is a community-driven research project actively seeking contributions across tools, skills, plugins, providers, multi-agent systems, testing, and documentation. Key contributor resources include CONTRIBUTING.md, CHANGELOG.md, and docs/SHOWCASE.md.

Licensing & Compatibility

Released under the MIT License, permitting commercial use and closed-source linking. It supports Anthropic-compatible API endpoints (including Moonshot/Kimi, Vertex, Bedrock) and is compatible with anthropics/skills and claude-code/plugins.

Limitations & Caveats

This is the initial open-source release (v0.1.0). Voice mode is not yet integrated into provider compatibility. Multi-agent coordination features like ClawTeam integration are listed on the roadmap.