Omnigent

Operator's take

You have Claude Code running on your laptop, a teammate on Codex, and a contractor on Cursor. None of them can see each other's sessions. Nobody knows what any of it cost until the bill lands. Omnigent is a layer that sits above the coding agent instead of replacing it: each agent gets wrapped in a uniform session so a server can watch it, cap it, and share it. It makes no individual agent smarter. It makes a group of them governable.

The control layer is the part worth your attention. Policies are stateful, so they track what a session has already done: cumulative LLM spend, tool calls made, risk accumulated. You can pause an agent every $100 it burns and make a human approve continuing. You can say that once an agent has installed a package from npm, a git push needs approval. That is a different thing from the allow/deny lists your coding agent already ships with, and it is the honest reason to look at this. Sharing runs on the same plumbing: a session gets a URL, a teammate opens it in a browser, and they can watch the stream, comment on files in the workspace, or send the agent commands.

The catch is maturity and cost of entry. Omnigent is alpha and says so on its own homepage. Installing it wants Python 3.12, Node 22 LTS and tmux on your machine, and there is no native Windows build (WSL2 only). That is a developer setup, not an operator one. The fully managed path (a Databricks-run server tied to your workspace identity, models through Foundation Model APIs and AI Gateway, agents in Databricks Sandboxes) is gated behind a workspace preview flag and a short list of AWS regions. It also drops the most interesting capability: custom Python policy functions do not run on Databricks-managed Omnigent, only the built-ins.

The fair comparison is Warp's Oz, which sells governance, routing and a team-wide view as a commercial control plane. Omnigent covers a lot of the same ground under Apache 2.0, so the price is your time rather than a seat fee. The awkward part is who built it: this comes from a data-platform company, and the managed route assumes you already run Databricks. If you don't, you are self-hosting an alpha project on Railway or Fly with a Postgres attached, and you own every upgrade. So the adoption question is narrow. If you have three or more people running coding agents and no idea what they're spending, Omnigent earns a weekend of your time. If you're one person with one agent, this is a layer you don't need yet.

What it's good at

• Stateful cost caps — policies track cumulative LLM spend across a whole session, so you can pause an agent every $100 and require a human to approve continuing.
• Harness swap in one line — a custom agent is a YAML file; change the harness value to move it from Claude Code to Codex while its prompts, tools, skills and policies stay exactly as they were.
• Live session sharing by URL — click Share, send the link, and a teammate watches the output stream, comments on files in the agent's workspace, forks the session, or sends commands to the running agent.
• Credentials the agent never sees — the bundled OS sandbox can intercept network egress and inject a token only on approved requests, so a GitHub secret is brokered rather than handed over.
• Local or cloud execution, one session — a runner executes on your own machine or on Modal, Daytona, E2B and other sandbox providers, so work keeps going after you shut the laptop.
• Real license, real traction — Apache 2.0, 9,766 stars and 1,516 forks as of 7 September 2026, with the last push that same day.

What it's not

• Not a coding agent — it drives Claude Code, Codex, Pi and Cursor rather than replacing them, and it adds close to nothing if you only ever use one agent.
• Not stable — alpha, three months old (repo created 11 June 2026), with 1,323 open issues against 9.7k stars. Expect breakage on upgrade.
• Not a no-code install — Python 3.12, Node 22 LTS and tmux are prerequisites, multi-user means deploying a server plus Postgres yourself, and Windows needs WSL2.
• Not fully itself on Databricks — the managed version blocks custom Python policy functions, and a Databricks Sandbox host always routes models through AI Gateway, so you can't bring your own model API keys.
• Not safe to share loosely — granting a teammate Edit on a session lets them run arbitrary code on whichever machine hosts it; the docs state this outright.
• Not first-class for every harness named — Claude Code, Codex and Pi are the terminal harnesses the announcement leads with, while Cursor, Copilot and Antigravity arrive as opt-in SDK-harness extras chosen at install time.