OpenClaw Enterprise is a free pilot control plane, not a 1.0

OpenClaw Enterprise is the company's attempt to make persistent agents acceptable to an IT team that would rather ban them. The announcement, on the OpenClaw blog, and a New Stack report opened for this piece agree on the shape: open source, self-hosted, free to run, and not a finished 1.0. They do not agree on every name in the sponsor list. This is a read of those two pages, not a setup guide and not a buying note.

TL;DR - OpenClaw Enterprise, called OCE in the post, is a control plane for persistent agents. The post says it is open source, vendor neutral, and free to run on your own infrastructure. - The same post says it is for internal pilot workloads, with a 1.0 release later this year. You can run it with docker-compose locally or on Kubernetes. - The New Stack says the architecture docs already mark some pieces as implemented and others as unfinished, including external gateway admission. That gap is not in the blog's headline.

What actually happened

The blog post says deployment of persistent agents is still limited almost a year after OpenClaw first launched. The reason it gives is blunt. Organizations want a stronger common standard for security, safety, and governance before they adopt agents. The default, it says, is that IT bans agent platforms like OpenClaw outright. OpenClaw Enterprise is the answer the post offers: multi-tenancy, hard security boundaries, and agent primitives that can be swapped, including the harness, the model, and the sandbox.

The post says the project started at OpenAI and was donated to the OpenClaw Foundation, where it is now an independent project, developed further with Red Hat and NVIDIA. It says OpenClaw Enterprise is being piloted inside companies like Red Hat and OpenAI, and that OpenAI is already running OpenClaw agents with full access to codebases and plugins. It says you can self-host today by cloning the repo. Local development is docker-compose. Internal deploy is Kubernetes. It says the software will always be free for any organization to use. The New Stack, reading the docs, says the Docker or Podman Compose option is currently a control-plane preview and cannot deploy agents through the control plane. The blog's sentence and that limit are both on the record. A local compose preview is not the same as an internal Kubernetes deploy.

Security language on the post is a list, not a score. Hard boundaries between trusted and untrusted workloads, sandboxing, LLM-based reviews, and fine-grained permissions. It says a reference architecture showing how those pieces fit is coming "in the coming weeks." That sentence is a schedule for a diagram, not a claim that the diagram already exists on the blog.

The blog byline is Kevin Lin, dated Sep. 29, 2026. The New Stack, read on October 2, 2026, calls the same announcement a Tuesday post by Kevin Lin, a member of technical staff at OpenAI leading the OCE work. September 29, 2026 was a Tuesday, so the day and the byline agree. Lin's line about IT banning agent platforms, as quoted there, matches the blog.

The New Stack adds a sponsor list the blog paragraph does not. It says the foundation's sponsors include OpenAI, Nvidia, Red Hat, and GitHub, plus other companies that committed resources. The blog names OpenAI, Red Hat, and NVIDIA as the development path. It does not, in the text that loaded, name GitHub. Both versions can sit side by side. One is the blog's origin story. The other is the newsroom's sponsor list.

What the docs say that the headline skips

The New Stack says it read the architecture documentation, not only the announcement. On that reading, the API, the console, a persistent worker, a PostgreSQL backend, and Kubernetes packaging are already implemented. External gateway admission, workload authentication back to the control plane, and some approaches to model authentication are not finished. The blog does not list those unfinished items. It says the project is early and that a reference architecture is still to come. The unfinished list is The New Stack's, from the docs. If you are deciding whether a pilot is even a pilot, that list matters more than the word "enterprise" in the title.

The control plane, which The New Stack abbreviates OCC, is described there as the admin surface: deploy agents, put them in isolated namespaces, manage configuration and credentials, set permissions, and keep a record of changes. Gateways receive messages. Harnesses handle turns, model calls, and tool execution. That split is the useful part. A control plane that records changes is not the same thing as an agent that is allowed to merge code.

The blog includes a testimonial about an internal agent called Androidclaw. The quoted user says it has context, git, GitHub, and logging, that it can find a pull request for a broken build, and that it answered one piece of feedback in "like 30 seconds" with a link to a SEV. The quote also says it traced a glitchy stream, published a pull request with video evidence, and merged it. That paragraph is on the company page. It is an anecdote with a casual time phrase, not a benchmark with a method. The New Stack does not use that anecdote in the passages that loaded. Do not treat "30 seconds" as a latency claim the newsroom confirmed.

What neither page is

Neither page is a 1.0 release note. The blog says the 1.0 is later this year and that today's build is for internal pilots. The New Stack says there is still distance before OpenClaw Enterprise is a finished enterprise platform. A reader who hears "enterprise" and pictures a support contract, a priced tier, or a hosted button will not find that product in these two pages. The post says it runs on your infrastructure and is free. Free to run is not free of the work of running it.

Neither page prints a model score, a price per token, or a customer count beyond the pilot names. Red Hat and OpenAI are the names on the blog. GitHub is extra, and only on The New Stack. NVIDIA is a collaborator on the blog and a sponsor on the newsroom list. The spelling on the blog is NVIDIA. The newsroom uses Nvidia. Same company, two styles. Not two companies.

People comparing agent launches can read the Cloudflare Clef note for a decision model with a published scoreboard, or the Robinhood Agents note for an in-app agent that places trades. OpenClaw Enterprise is neither of those. It is a control plane you host, and the pages say the 1.0 is still ahead.

What to do as a reader (not a trade)

This is not investment advice and not a suggestion to buy any token, stock, or seat tied to the names above. If you are an operator, read the pilot limit before the adjective. The blog says internal pilot workloads. The New Stack says parts of admission and authentication are unfinished. Clone OpenClaw Enterprise only if you are prepared to treat the build as early. Wait for the reference architecture the blog says is coming if you need the security picture in one diagram.

If you cite the origin, cite the blog: started at OpenAI, donated to the OpenClaw Foundation, further work with Red Hat and NVIDIA. If you cite GitHub as a sponsor, cite The New Stack, not the blog. If you cite the 30-second story, cite it as a quote on the company page, not as a measured result.

The pages are the OpenClaw announcement and The New Stack's report. Where the sponsor list differs, keep both versions. Where the docs list unfinished work, do not let the headline sand it off.