The license cost just went to zero

On September 29, the OpenClaw Foundation announced OpenClaw Enterprise (OCE), an open-source, vendor-neutral control plane for persistent AI agents. It is MIT-licensed. The foundation says OCE “will always be free for any organization to use.” You still pay for models, compute, storage, and the people who run it.

The backstory is what makes this news. OCE started inside OpenAI, was donated to the OpenClaw Foundation, and has since been developed with Red Hat and Nvidia. OpenAI and Red Hat are piloting it internally. The license file on GitHub still reads “Copyright (c) 2026 OpenAI.”

My view: “buy the model, rent the control plane” is now an outdated way to buy agents. When the governance layer is free and open source, the software stops being the expensive part. The expensive part is the discipline: identity, deny-by-default permissions, audit, and someone who owns all three.

What OCE actually is

It is not a model and it is not a chatbot. The project README puts it plainly: “Think of it as Kubernetes for agents.” At the center is the OpenClaw Control Plane (OCC). Administrators use it to deploy agents, separate them into isolated namespaces, manage configuration and credentials, set permissions, and keep a record of changes. The foundation lists multi-tenancy, hard boundaries between trusted and untrusted workloads, sandboxing, fine-grained permissions, and LLM-based review of agent actions. The harness, the model, and the sandbox are designed to be swappable. That is the layer I’ve argued buyers should diligence first, in enterprise AI harnesses beat model wars.

Red Hat and Nvidia work on the layers around it. Nvidia’s OpenShell is a separate open-source runtime that denies agents access by default, enforces policy outside the agent’s own process, and logs every allow and deny. Red Hat sponsors the foundation and has a dedicated OpenClaw team. It says it is bringing OpenClaw’s requirements into its Red Hat AI platform, including multi-tenant agent execution, identity-based tool filtering, and OpenShell capabilities on Kubernetes.

The foundation also explains why OCE exists: “the default stance of IT in most organizations is to ban agentic platforms like OpenClaw altogether.” OCE is its attempt to replace the ban with controls.

Why “free” matters to a CFO

Two reasons.

First, it resets vendor pricing. If a vendor charges a premium for agent governance (a registry, permissions, audit), you now have a free, open baseline to compare against. Commercial platforms still have a place. VentureBeat points to Runlayer, which raised a $30M Series A in June. Runlayer covers a wider scope: governing AI use across tools like Claude Code, Cursor, ChatGPT, and Codex, and finding unsanctioned “shadow AI.” But the vendor now has to show what you get beyond the free version.

Second, it separates two budget lines that usually get blended together. The model is a variable usage cost. The control layer is a fixed cost that you own. When they are blended, nobody owns the control layer. When they are separate, someone does.

A wall key cabinet holding a badge, padlock, safe, logbook, and stop button
A free control plane still needs overnight controls: identity, deny-by-default access, audit, and someone who owns the stop button.

What OCE isn’t yet

Don’t read “Enterprise” as “production-ready.” The foundation recommends OCE for internal pilot workloads, with a 1.0 release planned later this year. It says a reference architecture showing how the boundaries, sandboxing, LLM reviews, and permissions fit together is coming “in the coming weeks.” The New Stack, which read the architecture docs, reports that some pieces are still unfinished, including external gateway admission and workload authentication back to the control plane. Per the README, the simple Compose setup is only a control-plane preview and cannot deploy agents. Deploying agents requires Kubernetes.

One more detail. OCE is designed to be vendor-neutral, but the getting-started guide asks for an OpenAI API key for the default model when you deploy your first agent, and the project lead, Kevin Lin, works at OpenAI. Swapping models is supported. Test it rather than assume it works.

For most SMBs, the honest answer is that you won’t run OCE this quarter. Most 50-person companies don’t run Kubernetes or employ a platform engineer. But the checklist behind OCE applies whether you host it yourself, buy a commercial platform, or let a vendor run your agents.

The overnight checklist

OpenAI’s own example shows what’s at stake. It runs an internal agent called Androidclaw that works across company context, Git, GitHub, and logging. According to OpenAI’s RJ Marsan, it can find the pull request behind a broken build and fix it, and in some cases publish and merge the fix itself. The agent is useful because it has privileged access. That access is also the risk. Before any agent runs overnight in your company, put five things in one central place. It’s the same list I’d use for a spend and PII control plane for coding agents:

  1. Identity. Every agent gets its own identity and a named human owner. No shared service accounts.
  2. Deny by default. The agent starts with no access. It gets access one system and one action at a time: read, draft, write, or pay. OpenShell’s model is a good template.
  3. Credentials kept out of the agent. The platform hands out secrets as needed. Keys never go into prompts or into config files the agent can read.
  4. Audit you can read. Log every tool call and every permission change. Someone other than the builder reviews the log weekly.
  5. Kill switch and lifecycle. One place to pause an agent, rotate its credentials, and retire it when its owner leaves.

If you can’t say where each of those lives, the agent doesn’t run unattended. That’s the whole policy.

FAQ

Is OpenClaw Enterprise really free? The license is. OCE is MIT-licensed, and the foundation says it “will always be free for any organization to use.” You still pay for models, compute, storage, and the people who run it.

Is OCE ready for production? Not yet. The foundation recommends it for internal pilot workloads, with a 1.0 release planned later this year. The New Stack reports some pieces are still unfinished.

Do I need Kubernetes to run it? To deploy agents, yes. Per the README, the simple Compose setup is only a control-plane preview and cannot deploy agents.

Does OCE lock me into OpenAI models? It’s designed to be vendor-neutral, and swapping models is supported. The getting-started guide asks for an OpenAI API key for the default model, so test a swap rather than assume it works.

Bottom line

OpenClaw Enterprise doesn’t make agents safe. It makes the controls cheap enough that there’s no excuse to skip them. OpenAI started it, Red Hat and Nvidia are building around it, and it costs nothing to license. The discipline still has to come from you.

My advice: pick one internal agent and write down where each of the five controls lives. If you have a platform engineer, pilot OCE against that list before 1.0 ships. If you don’t, use the list to question whichever vendor runs your agents. If a vendor tells you governance is a premium add-on, ask what they do that the free version doesn’t. I may be early on how fast this matures. If your IT team is already running it, I’d like to hear what broke.

Sources

Written by

Ken Koo

Partner at Force Scaling

Partner at Force Scaling, helping CEOs and leadership teams turn ambition into operating discipline and results.

ADVERTISEMENT