omnigent

Purpose

Omnigent is an open-source AI agent framework and meta-orchestration layer (meta-harness), used to build a unified orchestration layer on top of Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and user-written agents. Its goal is to let users replace or combine different harnesses without rewriting, uniformly enforce policies and sandboxes, and support real-time collaboration from any device.

Main capabilities

• Multi-device session synchronization: start a session in the terminal, continue it in the browser, and then pick it up again on a phone; messages, subagents, terminals, and files stay synchronized. Provides a local Web UI (default port 6767) and a macOS desktop app (including system notifications, configurable sounds, and Dock badges).
• Multi-agent supervision: mix Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents defined in YAML in the same session, let one agent review another's work, or split tasks among agents that are each good at different things.
• Model selection: supports bringing your own API key, Claude/ChatGPT subscriptions, or any compatible gateway; default credentials are distinguished by agent, and the model can be switched in a session with the /model command.
• Real-time collaboration: share sessions so teammates can talk with agents, watch work in real time, co-drive on their machines, or fork the session to continue independently. Supports multi-user accounts, controlled via environment variables; on first run it outputs an administrator password, and one-time invitation links can be generated through the admin backend.
• Cloud sandbox execution: run sessions in one-off sandboxes such as Modal, Daytona, Blaxel, Islo, E2B, Gensee, CoreWeave, Kubernetes, OpenShell, Boxlite, microsandbox, Databricks, and others, without needing a laptop to stay online; supports startup by CLI or per-session provisioning by the server (managed hosts).
• Governance and policy: create policies to pause for approval before high-risk operations, limit spending, or restrict the tools an agent can access; policies can apply to the entire server, a single agent, or a single chat.
• Example agents: the repository includes examples such as Polly (a multi-agent coding coordinator that delegates coding subagents in parallel git worktrees and cross-reviews), Debby (Claude and GPT Dual-head brainstorming partner, supports /debate for mutual critique), Deep Research (a deep research agent with citations and cross-verification, using an MCP search server).

How to Use

• Installation: Install Omnigent and its dependencies with a single command. Requires Python 3.12+, uv, git; the programming harness CLI requires Node.js 22 LTS or newer and npm, the Web UI requires pnpm; the native terminal wrapper requires tmux; on Linux, bubblewrap (bwrap) is required for OS sandboxing, while macOS uses the built-in seatbelt. Optional integrations can be installed via extras: model providers (databricks, bedrock, vertex), sandbox providers (modal, daytona, blaxel, boxlite, microsandbox, cwsandbox, e2b, openshell, kubernetes), SDK harnesses (antigravity, copilot, cursor, agents-sdk), storage and memory (s3, hindsight). It can also be installed via Homebrew or directly from the repository.
• Launch: After running, Omnigent will choose a model together with the user and start a session in the terminal, while also starting the local Web UI. After installation, two equivalent commands, omnigent and the shorter omni, will be on PATH. You can launch a specific agent runtime (e.g., omnigent agy), or use the browser approach: start the local server with one command and register this machine as a host, open the printed server URL, click New Chat, and select your own machine.
• Model configuration: Add, set as default, or remove credentials, grouped by agent; defaults are differentiated by agent, so a Claude default and a Codex default can coexist.
• Deploying a server: After running on a server with a stable URL, sessions can be accessed from anywhere (including a phone); the Web UI is built for mobile. Supports one-click deployment with docker compose, Render, Railway, as well as Fly.io, Hugging Face Spaces, Modal, Cloudflare (serverless, scales to zero), Databricks Apps, etc.; it can also be accessed from your own laptop via a Cloudflare quick tunnel (public) or Tailscale (private). Within your own network, you can directly open the local LAN address on your phone.
• Stopping and status: Use omnigent server status to check status, and omnigent stop to stop everything.
• Upgrade and uninstall: When a new version is released, Omnigent will show a one-line notice (once per version), and you can update with omni upgrade; uninstalling can clean up the CLI and the installer-managed PATH entries, --purge can additionally remove the ~/.omnigent state, and --purge-workspace can remove the ~/omnigent workspace.

Use Cases

Suitable for developers who need orchestration and supervision across multiple coding agents, teams that need access from multiple devices and to share live sessions across the team, and organizations that need to run agents in cloud sandboxes and impose approval, spending, and tool access policies on agent behavior. It can also be used for automation: the Automations page can run agents periodically on a schedule.

Limitations

• On Windows it runs natively in a degraded mode: the installation bootstrap script is POSIX-only and requires installing directly with uv. On Windows, omnigent server, the Web UI, and SDK-based harnesses are available, and agents run under a Windows Job Object for process tree isolation; however, the native tmux/PTY terminal wrapper is unavailable, bwrap/seatbelt filesystem and network sandboxing and the L7 egress proxy are unavailable, and the Job Object backend only contains the process tree and enforces resource limits, without isolating the filesystem or network.
• Devin CLI and Kiro CLI are optional integrations that require separate installation and login; Devin tool approvals appear as Chat approval cards, while Kiro tool approvals are answered in the embedded terminal.
• The XAI-related built-in ACP line does not have its own env_passthrough, and XAI_API_KEY is not in the credential allowlist from host to runner; exporting it directly will not reach the agent, so it must be passed explicitly or an acp agent declaring passthrough must be configured.
• omnigent agy requires agy 1.1.13 or newer; setting GWhen using EMINI_API_KEY, direct Gemini API authentication takes precedence over the OAuth login saved by agy.
• Title generation has length limits: by default, generated titles do not exceed 100 characters, custom titles are required to be at most 200 characters, and overly long ones will be truncated; manually specified titles are also limited to 200 characters.
• Multi-user accounts require the environment variable OMNIGENT_AUTH_ENABLED to be enabled; Docker deployments have it enabled by default; registration is invitation-based, teammates need to be able to access the server, a local server is only reachable on its own network, and offline users need to deploy an always-on host.
• The full menu of deployment targets, database options, sandbox settings, and branding/white-label related content are located in deploy/README.md.
• Materials such as charges, pricing, and specific effects were not provided.