OpenHarness

OpenHarness is an open-source Python framework from HKUDS that delivers core lightweight agent infrastructure: tool-use, skills, memory, and multi-agent coordination. It ships alongside ohmo, a personal AI agent built on top of the harness that integrates with Feishu, Slack, Telegram, and Discord. Released under the MIT license in April 2026, the project had accumulated over 12,800 GitHub stars and 2,100 forks as of its About page data.

What It Is

OpenHarness implements what it calls the Agent Harness pattern — the complete infrastructure that wraps around an LLM to make it a functional agent. The model provides intelligence; the harness provides tools, memory, safety boundaries, and coordination primitives. The framework is structured around 10 subsystems: engine (agent loop), tools, skills, plugins, permissions, hooks, commands, MCP client, memory, and a multi-agent coordinator. It is designed for researchers, builders, and the open-source community who want to understand, experiment with, and extend production-style AI agent architecture.

Harness Architecture and Agent Loop

The core of OpenHarness is a streaming agent loop: the model receives a prompt, streams a response, executes tool calls through a permission-and-hook pipeline, appends results to the message history, and loops until the model stops requesting tools. The harness handles safety, observability, and efficiency while the model decides what to do.

Key subsystems include:

• 43+ tools across file I/O, shell, web search/fetch, MCP, notebook editing, subagent spawning, task management, scheduling, and meta operations
• Skills system — on-demand Markdown knowledge files loaded only when needed, compatible with the anthropics/skills directory layout
• Plugin system — compatible with claude-code plugins; tested with 12 official plugins covering commit workflows, security hooks, code review agents, and PR tooling
• Permissions — multi-level modes (Default, Auto, Plan Mode) with path-level rules, command deny lists, and PreToolUse/PostToolUse lifecycle hooks
• React/Ink TUI — interactive terminal UI with command picker, permission dialogs, session resume, and animated feedback

Provider Compatibility

OpenHarness treats LLM providers as named workflow profiles. Built-in workflows cover:

• Anthropic-Compatible API (Claude official, Moonshot/Kimi, GLM, MiniMax)
• Claude Subscription bridge via local credentials
• OpenAI-Compatible API (OpenAI, OpenRouter, DeepSeek, DashScope, GitHub Models, SiliconFlow, NVIDIA NIM, Google Gemini, Groq, Ollama)
• Codex Subscription bridge
• GitHub Copilot via OAuth device flow — no API key needed

Profiles are scoped so different Anthropic-compatible or OpenAI-compatible endpoints can hold separate credentials. Local models via Ollama are supported through the OpenAI-compatible endpoint.

ohmo Personal Agent

ohmo is a packaged personal-agent app built on OpenHarness with its own ~/.ohmo/ workspace, gateway, and channel configuration. After running ohmo init and ohmo config, users start a gateway that connects to Telegram, Slack, Discord, or Feishu. From those chat apps, ohmo can fork branches, write code, run tests, and open PRs autonomously. The project states that ohmo runs on an existing Claude Code or Codex subscription with no extra API key required.

Update: v0.1.9 — Skill Workflows and Provider Key Updates

The latest release is v0.1.9 (published 2026-05-07), titled "Skill Workflows and Provider Key Updates." The changelog documents rapid iteration since the initial v0.1.0 open-source release on 2026-04-01:

• v0.1.7 (2026-04-18): Install script improvements, React TUI Shift+Enter newline support, quieter Windows spinner
• v0.1.6 (2026-04-10): Auto-Compaction for multi-day sessions, headless worker mode for subprocess teammates, full Markdown rendering in TUI
• v0.1.5 (2026-04-08): MCP HTTP transport, auto-reconnect, JSON Schema inference for MCP tool inputs
• v0.1.4 (2026-04-08): Native Moonshot/Kimi provider, auth overhaul, sensitive-path protection, EIO crash recovery
• v0.1.2 (2026-04-06): oh setup workflow wizard, ohmo packaged app launch

An unreleased --dry-run feature is documented that previews resolved runtime settings, auth state, skills, commands, tools, and MCP servers without executing the model or tools, returning a ready/warning/blocked readiness verdict.

Testing and Extensibility

The repository reports 114 passing unit and integration tests plus multiple E2E suites covering CLI flags, harness features, React TUI interactions, and real skills/plugins. Custom tools are added by subclassing BaseTool with a Pydantic input model. Custom skills are Markdown files placed in ~/.openharness/skills/. Plugins follow the .claude-plugin/plugin.json convention and can include commands, hooks, and agents.