OpenHarness
An open-source AI agent framework for macOS, Windows and Linux, with Ollama support, persistent memory and a personal agent accessible through chat apps.
OpenHarness is an MIT-licensed Python framework for developers and researchers building AI agents on their own machines. It runs on macOS, Linux and native Windows, and includes ohmo, a personal assistant that can work on coding tasks through Feishu, Slack, Telegram or Discord.
The framework connects to local LLMs through Ollama or to cloud providers through OpenAI-compatible and Anthropic-compatible APIs. Supported services include Claude, DeepSeek, Kimi and Google Gemini. The agent software runs on your machine; choosing a cloud backend sends model requests to that service. Ohmo can also use an existing Claude Code or Codex subscription without a separate API key, and OpenHarness supports GitHub Copilot authentication.
Agents can read and edit files, execute shell commands, search code and the web, and use MCP tools. Ohmo can create branches, write code, run tests and open pull requests. For larger jobs, OpenHarness supports delegated subagents, shared task management and background work.
Persistent memory and resumable sessions help retain work across conversations. Automatic context compression preserves task state and channel logs during long sessions. Developers can extend agent behavior with Markdown skills and Claude Code-compatible plugins, including custom hooks and agents.
Permission controls let you restrict file paths and commands and require interactive approval for tool actions. The terminal interface supports Markdown, while script integrations can receive JSON results or streamed events.
