The whole arc branches.
Not the front end. Not the database. A branch of your code gets a complete, isolated copy of your whole arc. An arc is compute, Postgres, cache, objects, and its own URL — carrying real production data with the personal details scrubbed on the way in.
Measured on real hardware: a branch of a 30 MB arc costs 88 KB and is answering queries in under a second. Pre-launch, building with design partners.
Jurisdiction, not location.
An American provider answers to American courts wherever its disks sit. The EU's proposed Cloud and AI Development Act conditions public procurement on EU ownership, and no amount of Frankfurt fixes that.
- European bare metal, European company, European certificate authority.
- Our own edge rather than a US CDN.
- This page loads nothing from anyone: no fonts, no CDN, no analytics, no trackers.
# arc.yaml
arc: acme
applications:
api:
kind: web-service
image: registry.omniarc.eu/acme/api:2.1
port: 8080
resources: { cpu: "1", memory: "1Gi" }
scale: { min: 0, max: 8, idle_timeout_secs: 120 }
main:
kind: database
version: "16"
contains_pii: true
resources: { cpu: "2", memory: "4Gi" }
anonymize:
- application: main
rules:
- { table: users, column: email, transform: fake-email }
One file, two surfaces, no platform team.
Everything OmniArc can do is reachable from the HTTP API and from
omniarc, the CLI that is a client of it. No Terraform. No dashboard
you have to learn. No second infrastructure language.
An idle branch bills zero.
A branch costs 88 KB to create and bills only for what diverges afterwards. Asleep, it costs nothing but its bytes. Here is the same workload — one production stack and twenty preview branches — priced on the platforms you are probably using today.
Built for the way you already work.
- An agent working on a branch has a complete, anonymized copy of production it
can destroy for the cost of one
zfs destroy. Blast radius is why agents are kept away from infrastructure; branching removes it. - The same verbs a human gets. An MCP server over the same API ships with the API, not after it. (planned)
- Nonsense fails at parse time, with every error at once and a reason for each — which is what makes an agent's edit-and-validate loop converge.
Measured on real hardware.
The two wake times were measured on a small, low-power test box — not production hardware. Read them as "fast enough to disappear," not as a number to hold us to.
Nine of twelve feasibility claims are measured. The three that are not are blocked on hardware we don't yet have, and are listed as unanswered. Full numbers are in our internal findings write-up — not yet public.