Omnigent describes itself as a **meta-harness** — an orchestration layer that sits above coding agents rather than being one. It is Apache-2.0, written mostly in Python, and has gathered roughly **8,009 stars and 1,190 forks** since the repository opened on **11 June 2026**, one of the steeper adoption curves of the year. The project ships releases on a weekly-ish cadence (**v0.7.0** on 27 July 2026, v0.6.0 on 21 July) and still carries an explicit **alpha** status badge. ## The Problem It Targets The agent-harness market fragmented fast. Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi each have their own session model, configuration format, approval flow, and sandboxing story. Committing to one means your prompts, policies, and team workflows are written against that vendor's surface — and switching later means rewriting them. Omnigent's proposition is to make the harness a swappable component. You define agents once (custom agents are plain YAML), and you can **mix multiple harnesses inside a single session** — asking one agent to review another's work, or splitting a task across agents with different strengths. The governance and collaboration layer lives in Omnigent, so it survives a harness swap. ## Sessions That Follow You The most-cited feature is device continuity. A session started in the terminal continues in the browser and can be picked up on a phone, with messages, sub-agents, terminals, and files staying in sync. The web UI is explicitly built for mobile, and there is a native macOS desktop app alongside it. On a local network no deployment is needed at all — opening the machine's LAN address (`http://192.168.x.x:6767`) on a phone is enough. ## Credentials and Models Omnigent treats four credential kinds as first-class, which matters for teams that mix billing models: | Kind | What it is | |---|---| | API key | First-party vendor keys (Anthropic, OpenAI, and similar) | | Subscription | A Claude Pro/Max or ChatGPT plan, via the official `claude` / `codex` CLIs | | Gateway | Any OpenAI- or Anthropic-compatible `base_url` — OpenRouter, LiteLLM, Ollama, vLLM, Azure | | Databricks | A Databricks workspace profile, via the `databricks` extra | Defaults are per-agent, so a Claude default and a Codex default coexist without conflict, and `/model` switches mid-session. The docs are precise about a gotcha that bites people in practice: pointing Claude Code at OpenRouter needs the Anthropic-compatible `…/api` endpoint, while Codex needs `…/api/v1`. ## Sandboxing and Governance This is where Omnigent diverges from being a convenience wrapper. **Policies** evaluate every action an agent attempts — shell commands, file edits, token spend — and allow it, block it, or pause for human approval. Policies can be scoped to the whole server, a single agent, or one chat. OS-level isolation backs this up. On Linux, each agent terminal is wrapped in **bubblewrap** (`bwrap`), and the README is blunt that this is mandatory: a missing `bwrap` binary makes those terminals fail to start outright rather than falling back to an unsandboxed mode. macOS uses the built-in **seatbelt** sandbox. Sessions can also run entirely in disposable cloud sandboxes — **Modal, Daytona, Islo, E2B, CoreWeave, Kubernetes, OpenShell, Boxlite, and Databricks** are supported — launched from the CLI or provisioned per-session by the server as *managed hosts*, so no laptop needs to stay online. ## Team Collaboration Multi-user mode is one environment variable (`OMNIGENT_AUTH_ENABLED=1`, on by default in the Docker deploy). Signup is invite-only through single-use links generated in the admin UI, with no email server required, and `OMNIGENT_OIDC_ISSUER` wires up **Google, GitHub, Okta, or Microsoft** logins for teams that already have SSO. Three collaboration modes are distinct and thoughtfully separated: - **Share** — teammates watch your agent work live and chat with it - **Co-drive** — `omnigent attach <session_id>` lets a teammate's messages execute on *your* machine, for pairing or handing the keyboard to a domain expert mid-investigation - **Fork** — `omnigent run --fork <session_id>` clones the conversation onto their own machine to continue independently Shared sessions label model-visible messages with `[account]:` attribution by default so the agent knows who it is talking to. ## Deployment One `docker compose up` covers a VPS or home server; **Render** and **Railway** are one-click; **Fly.io, Hugging Face Spaces, Modal, Cloudflare** (serverless, scale-to-zero), and **Databricks Apps** (backed by Lakebase Postgres and Unity Catalog Volumes) are documented. A **Cloudflare quick tunnel** or **Tailscale** exposes a laptop-hosted server without deploying anything at all. Install is a single curl-to-shell script, or `uv tool install omnigent` / `brew install omnigent-ai/tap/omnigent` for the manual route. ## Caveats The alpha badge is not decorative — there are **815 open issues**, and the commit log shows active churn in the web UI and session-management internals as recently as 1 August 2026. The prerequisite list is long for a tool positioned as a convenience layer: Python 3.12+, Node.js 22 LTS with npm and pnpm, `uv`, `git`, and `tmux` for the native terminal wrappers. **Windows runs in an explicitly degraded mode.** The POSIX-only bootstrap does not apply, the native tmux/PTY terminal wrappers are unavailable, and — the substantive limitation — the Windows Job Object backend contains the process tree and enforces resource limits but **does not isolate the filesystem or network**. Teams that adopted Omnigent specifically for its sandboxing should treat Linux or macOS as the supported surface and route Windows users through WSL. ## Verdict Omnigent is the most complete attempt yet at making the coding-agent harness a replaceable part instead of a lock-in point, and it pairs that with the governance layer — approval gates, spend caps, tool allowlists, real OS sandboxing — that most agent tooling leaves as an exercise for the reader. Teams running several agents across several people, especially those who need an audit story before letting agents touch a repository, will find nothing else in the open-source landscape covering this much ground. Solo developers happy inside one harness are paying a real setup tax for flexibility they may not use, and everyone should read "alpha" literally before pointing it at production.