Omnigent

Omnigent is a meta-harness: one orchestration layer over the coding agents you already run. Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and agents you define yourself in YAML share a single session model, so you can mix them in one conversation, ask one agent to review another's work, or move a task between harnesses without rewriting it.

A session is not tied to a terminal. Starting omnigent runs the agent locally and also serves a web interface on port 6767 that shows the same session in a browser, on a phone on your network, or in a macOS desktop app; messages, sub-agents, terminals, and files stay in sync across all of them. Deploying the server to a host with a stable URL — Docker Compose, Render, Railway, Fly.io, Hugging Face Spaces, Modal, Cloudflare, or Databricks Apps — makes sessions reachable from anywhere, and a Cloudflare quick tunnel or Tailscale exposes a laptop without a deploy.

Credentials come in four kinds: a first-party vendor API key, a Claude or ChatGPT subscription used through the official CLIs, any OpenAI- or Anthropic-compatible gateway such as OpenRouter, LiteLLM, Ollama, vLLM, or Azure, and a Databricks workspace profile. Defaults are set per agent, and /model switches mid-session.

Policies decide what an agent may do. They check each action and allow it, block it, or pause for your approval, and they stack across three levels — server-wide, per-agent, and per-session — with the stricter session rules checked first. Built-in policies cover shell and file-write approval, a cap on tool calls per session, and spend budgets with soft warning thresholds. Agent terminals are wrapped in an OS sandbox — bubblewrap on Linux, seatbelt on macOS — and sessions can instead run in disposable cloud sandboxes from Modal, Daytona, Blaxel, Islo, E2B, CoreWeave, Kubernetes, OpenShell, Boxlite, or Databricks.

Multi-user accounts are enabled with one environment variable and are on by default in the Docker deploy. Signup is invite-only through single-use links, optional OIDC lets a team sign in with existing Google, GitHub, Okta, or Microsoft logins, and a live session can be shared for teammates to watch and chat, co-driven so a teammate's messages execute on your machine, or forked to continue independently.

An agent is a short YAML file naming a prompt, an executor harness, and tools — local Python functions, MCP servers, and sub-agents a supervisor delegates to. Two examples ship with the project: Polly, a multi-agent orchestrator that delegates coding work to sub-agents in parallel git worktrees and routes each diff to a reviewer from a different vendor, and Debby, a brainstorming partner running one Claude head and one GPT head. The project describes itself as alpha, requires Python 3.12 or newer, and runs in a degraded mode on native Windows without filesystem or network sandboxing.

Features

• Harness switching: Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, Antigravity, and custom YAML agents behind one interface
• Mixed sessions: several agents work in one session, including cross-vendor review of each other's output
• Device-independent sessions: terminal, browser, phone, and a macOS desktop app show the same live session
• Credential kinds: vendor API keys, Claude or ChatGPT subscriptions, OpenAI- or Anthropic-compatible gateways, and Databricks workspaces
• Layered policies: server, agent, and session rules that allow, block, or pause an action, with spend caps and tool-call limits
• OS sandboxing: bubblewrap on Linux and seatbelt on macOS wrap each agent terminal, with an L7 egress proxy
• Cloud sandboxes: sessions run in disposable Modal, Daytona, Blaxel, Islo, E2B, CoreWeave, Kubernetes, OpenShell, Boxlite, or Databricks environments
• Team accounts: invite-only signup, optional OIDC sign-in, and role-separated multi-user servers
• Live collaboration: share a session to watch and chat, co-attach to drive it on the owner's machine, or fork it