An open chassis for agents that won't ghost you mid-task
OpenHarness is a Python framework that gives LLMs the scaffolding—tools, memory, permissions, multi-agent coordination—to actually finish multi-day jobs without human hand-holding.

What it does
OpenHarness wraps raw LLMs in the infrastructure they need to act: 43+ built-in tools (file, shell, search, web, MCP), persistent MEMORY.md storage, auto-compaction that preserves state across long sessions, and permission hooks so your agent doesn’t rm -rf / on a whim. It ships with oh, a CLI entry point, and ohmo, a personal agent that plugs into Slack, Telegram, Discord, or Feishu and can fork branches, write code, run tests, and open PRs while you sleep.
The interesting bit
The ohmo personal agent runs on your existing Claude Code or Codex subscription—no extra API key. That’s a pragmatic hack: it piggybacks on subscriptions developers already have rather than inventing a new billing layer. The harness also treats context compression as a first-class problem, with auto-compaction designed to keep multi-day sessions alive without manual clear or compact rituals.
Key highlights
- 43+ tools with on-demand skill loading from
.mdfiles and a plugin ecosystem (skills + hooks + agents) - Multi-level permission system with path-level rules, pre/post tool hooks, and interactive approval dialogs
- Subagent spawning, team registry, and background task lifecycle for multi-agent swarms
- MCP protocol support with HTTP transport, auto-reconnect, and JSON Schema inference for tool inputs
- React Ink TUI with markdown rendering, plus non-interactive JSON output for scripting
- Native Windows support (with
openhalias to avoid PowerShell’s built-inOut-Hostconflict)
Caveats
- ClawTeam integration is on the roadmap, not shipped
- Windows TUI has needed repeated stability fixes; the changelog mentions “quieter” spinner frames and “reduced flashing” as ongoing work
- v0.1.7 is still early; the dry-run readiness verdict (
ready/warning/blocked) suggests the authors are actively discovering failure modes
Verdict
Worth a look if you’re building agents that need to run unsupervised for hours or days and you want inspectable Python infrastructure rather than a black-box SaaS. Skip it if you just need a quick ChatGPT wrapper—this is for people who want to own the harness, not just rent the horse.