Omnigent — Open-Source Meta-Harness

GitHub: github.com/omnigent-ai/omnigent Stars: 7,235 | Language: Python | License: Apache-2.0 Active: Very high (1380 commits, ~100/week)

Core Positioning

Open-source meta-harness providing a unified orchestration layer over Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents — switch or combine harnesses without rewriting, enforce policies and sandboxing, collaborate from any device.

Key Features

1. Multi-Agent Unified Orchestration — run Claude Code, Codex, Cursor, Hermes together in the same session; one agent reviews another’s work
2. Cross-Device Sync — terminal to browser to phone; sessions follow you (Kiro protocol)
3. Cloud Sandboxes — Modal, Daytona, E2B, CoreWeave, Kubernetes, Databricks, 8+ providers
4. Policy Governance — pause-for-approval before risky actions, spend caps, tool whitelists at server/agent/session granularity
5. Desktop App — native macOS + Web UI (${LOCALIP}:6767) + mobile
6. Modular Harness — SDK-based (YAML) + native tmux terminal wrappers
7. Sandboxing — bubblewrap (Linux), seatbelt (macOS), Job Object (Windows)
8. Collaboration — share sessions, teammates co-drive or fork

Why This Matters

Omnigent is the first “agent orchestration layer” — not another agent framework, but a management layer on top of existing agents. It solves the real problem of switching between Claude Code, Codex, Cursor without shared sessions, unified policies, or cross-agent collaboration. Complements Agent Orchestrator (which focuses on parallel coding supervision). Represents Agent infra evolution from tools-layer -> framework-layer -> orchestration-layer.