Cloudroom Measured: An Open Kernel, One Author, Four Weeks Old

On 14 October 2026 David Ondrej posted a sixty-minute video on his YouTube channel, which carries more than 400,000 subscribers: Cloudroom. The claims inside: every agent gets its own isolated cloud VM, his is the world's first AI coding software with cloud agents already on a free plan, and everything is fully open. The presenter is the founder himself, which makes the video a sales pitch. For that reason we measured the repo, the licence file, the website, the pricing page and the security page ourselves on 14 October 2026 at around 18:30 UTC. We executed nothing.

What the repository says

The kernel (cloudroom-core) was created on 17 September 2026, the desktop app (cloudroom-gui) on 19 September, both at roughly four weeks of build time as counted on 14 October. The LICENSE file contains the full Apache-2.0 text, 11,357 bytes, no badge trick: the API reports Apache-2.0 for both repos, and the GUI additionally carries a NOTICE with the MIT notice of the upstream project. The Rust share from the video ("78 percent") is almost exact: 921,718 of 1,181,706 counted bytes are Rust, which is 78.0 percent.

The release pace is fast: three releases on 13 October alone, core-v0.1.57 at 04:22 UTC on the morning of 14 October. The other side of that tempo: the contributor list of both repos holds exactly one person, the founder. The kernel repo held 42 commits on 14 Oct, the GUI repo 48; by 19 Oct they stand at 50 and 55. The kernel's issue tracker reported a single open item on 14 Oct, a feature idea from an outside user; two items stand open on 19 Oct. Nine subscribers watch the kernel now, the GUI repo had none on 14 Oct.

The author's own profile deserves a quick look too: the GitHub account dates from April 2023, the profile lists 28 public repositories, a fair number of them created on a single day and never touched again. The two Cloudroom repos, by contrast, were being pushed to right up through the early hours of 14 October, which separates them from the usual dump. Maturity, meanwhile, is something the README states itself: "This is pre-release software."

Stars against watchers

Stars and forks are snapshots: on 14 Oct, right after the video, they stood at 524 and 96; on 19 Oct at 898 and 193. Watchers are the people subscribed to updates; authors is the count of humans with commits. (Quelle: GitHub API, own measurements 14 Oct and 19 Oct 2026)

The video named 320 stars at recording time; we measured 524 on 14 Oct and 898 on re-measurement on 19 Oct. Between advertising and use there are naturally worlds in a freshly launched project. The watcher column is the only one of these figures that will still say something two weeks after the noise has faded.

The README is more honest than the video

The open kernel is a Rust service on your own Linux machine: an Ubuntu 24.04 setup in one script, PostgreSQL for the chat history, Tailscale for access on the road, the API listening on localhost only by default and demanding a 256-bit secret. Agents run as a separate Linux user without sudo, and they cannot read the kernel's token, database password or history files.

Next to that, the README carries a sentence the video does not show: "Agents run without approval prompts, so give each trusted user their own machine." Self-hosted therefore means isolation by operating-system user on one machine intended for one trusted person. The advertised VM per agent is an architecture property of the hosted product. And hosting the kernel alone brings a grant without tenancy: one script sets up Ubuntu, two SQL files you apply yourself, because the kernel never runs its own migrations. The site names the supported harnesses in two sets: nine locally (Claude Code, Codex, Pi, Cursor, OpenCode, Grok, fx, Hermes, Devin), six in the cloud.

Pricing page, security page, one blank

Free ($0) means 100 cloud agents per 30 days, 25 hours of cloud compute, 60 days of storage, unlimited locally, with an admission queue. Worked out, that is roughly 15 minutes of compute per agent, which leaves open how long a real task may actually breathe in there. Pro costs $50 a month (1,000 agents, 300-plus hours, twice as powerful sandboxes, one year of storage), Ultra $500 (10,000 agents, five years of storage), Enterprise comes later. All prices are stated tax-inclusive on the page.

The security page is unusually open: it lists the measures (sandboxes never shared, a token valid only for its own sandbox, direct connections to your model providers with Cloudroom never in between, secrets entered through a form that keeps them out of chat and logs, a Command Guard against a few catastrophic commands), and, unusually, it lists its own limits in a section titled "Know the limits". Three sentences there matter before any deployment: agents run without approval prompts with admin rights inside their sandbox and can read the keys stored there. Subagents share the lead agent's sandbox. And with Mac access switched on, a cloud agent can "run commands on your Mac as you".

The blank next to that list is larger than the list itself: no hosting provider and no country is named anywhere on the home, pricing, security or docs pages. For companies with data custody duties that is the first checkpoint, and it stays unanswered.

What we could not measure

The video demonstrates magnitudes we cannot see from outside: 38 concurrent agents, 201 open threads, a hundred subagents launched from one prompt, plus a referral rate of 11 to 15 percent. Those are vendor statements on launch day; we neither believe nor doubt them, we only note that there is no independent trace of them yet. One follow-up question we could settle: the free plan sits in an admission queue, both paid plans give instant access, and in the video the founder promises entry "in waves" depending on capacity. Anyone planning around a milestone plans against those three sentences. In passing: every documentation page also exists as plain Markdown, with an llms.txt alongside. For companies that build assistants or write for them, that is the most courteous entry of this whole research; measured while crawling the docs page.

The superlative in the title

The OpenAI docs for Codex Cloud have stated for some time that each task has its own workspace. Isolated cloud workspaces per agent or per task are therefore not new, only closed and locked to one harness. What Cloudroom genuinely adds is the combination: an open self-hostable kernel, a free choice of harness and model provider through your own subscriptions, and a hosted free plan. The video's headline claim is founder marketing. And even the foundation is not his own: the GUI is, quoting the README, "Built on BB", the MIT-licensed project get-bb/bb with 4,464 stars, 629 open issues (read on 19 Oct 2026) and 47 MB of TypeScript.

What this means for us

We run our agents in our own runtime on our own hardware, with our own harness and our own knowledge stores; the kernel overlaps with what we already operate and would be too small a replacement. The foundation adds to that math: under the GUI sit BB's 629 open issues (as of 19 Oct), and every upgrade upstream reaches into your own tooling. The hosted variant means handing source code and subscription logins to a startup that does not name where it stores them; that is a policy decision for the management. What the publication is worth: a security page with honest limits, a README that names the limits of self-hosted isolation itself, and a template for how founder videos become measurable. We keep it as a watch item; in four weeks there will be more to see.