Omnigent

What it does

• One CLI, many harnesses: omnigent claude / codex / cursor / opencode / hermes / pi launch native sessions; Grok Build and Devin are built in via omnigent run --harness (they use their own vendor CLIs to log in)
• Agents are short YAML files: prompt, tools (local Python functions, MCP servers, sub-agents a supervisor delegates to), harness choice — agents can even author other agents from a description
• Policies decide what an agent may do: pause for approval before risky actions, cap spend (cost_budget with soft warning thresholds), or limit tool calls — stacking server-wide, per-agent, and per-session
• Sessions follow you: start in the terminal, continue in the local web UI (localhost:6767), pick up on your phone, or share so teammates can watch and co-drive
• Cloud sandboxes are providers: Modal, Daytona, E2B, Kubernetes, Databricks, Blaxel, CoreWeave, microsandbox and more, launched per session or provisioned as managed hosts

When to use it

For teams that refuse to bet the workflow on a single CLI: keep Claude Code and Codex side by side in one session, let a Codex subagent review Claude's patch, and cap the spend with a policy instead of a spreadsheet. The YAML agent format also fits platform teams distributing standard agents (with pinned tools and policies) to the whole org.

Getting started

One command installs everything:

curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | sh

Or manually (Python 3.12+):

uv tool install omnigent
# or:
brew install omnigent-ai/tap/omnigent

The install puts both `omnigent` and the shorter `omni` on your PATH. On first run it picks up credentials already in your environment (ANTHROPIC_API_KEY / OPENAI_API_KEY, or a logged-in claude / codex CLI).

Things to know

Alpha: expect interface churn — pin versions. Linux needs bubblewrap for the native terminal wrappers' mandatory OS sandbox; Windows runs natively only in degraded mode (SDK harnesses and web UI, no filesystem/network isolation). Native agy does not accept OpenAI-compatible endpoints, and the Antigravity SDK extra conflicts with Databricks dependencies.