OpenHarness is the scaffolding that turns LLMs into actual coworkers

What it does

OpenHarness is a Python framework that wraps around LLMs to make them functional agents. It handles the unglamorous but essential plumbing: executing shell commands and file operations, compressing context when conversations grow too long, persisting memory across sessions via MEMORY.md, and enforcing path-level permissions so an agent can’t wipe the wrong directory. The project also ships ohmo, a personal agent that plugs into Slack, Telegram, Discord, or Feishu and can write code, run tests, and open pull requests on its own.

The interesting bit

The clever angle is that ohmo piggybacks on your existing Claude Code or Codex subscription—no extra API key required—and the harness includes a --dry-run mode that previews exactly what tools, skills, and auth state an agent will use before it touches anything. It also treats context compression as a first-class concern, preserving task state across auto-compaction so agents can theoretically run multi-day sessions without manual intervention.

Key highlights

• 43+ built-in tools covering file system, shell, web search, and MCP servers, with on-demand skill loading from markdown files
• Multi-agent swarm coordination with subagent spawning, background task lifecycle, and team registries
• Governance layer with multi-level permission modes, pre/post-tool hooks, and interactive approval dialogs
• Native support for multiple providers (Claude, OpenAI, Codex, Moonshot/Kimi, GLM, MiniMax, NVIDIA NIM) with profile-scoped credential management
• React Ink-based TUI with Markdown rendering, plus a non-interactive JSON mode for scripting

Caveats

• Windows users must use openh instead of oh because the command conflicts with PowerShell’s built-in Out-Host alias
• Some features like ClawTeam integration remain on the roadmap rather than in the current codebase
• At v0.1.7, the project is young enough that recent churn around MCP transport and auth credential priority suggests the surface is still shifting

Verdict

Researchers and builders who want to understand or extend production agent infrastructure will find this more useful than casual users looking for a drop-in chatbot. If you already pay for Claude Code or Codex and want an open-source layer that actually uses those subscriptions to do work in the background, this is built for you.