Perceo has one hard constraint: the machine that holds your code, your VMs, and
your notes is a machine you control. Everything else follows from that.
What runs where
There is no Perceo account, no Perceo-hosted control plane, and no telemetry
endpoint. If you point the model calls at a local model, nothing leaves at all.
Credentials stay where you put them
Archductor does not store or proxy credentials. It drives the CLIs you have
already authenticated:
gh auth login for pull requests, checks, comments, and merges- your existing
codex and claude local authentication for agent sessions
openssh for SSH repository access
Archivum reads its keys from a local .env written during setup. Archfleet uses
your hypervisor’s own access model.
Local-first is not the same as offline. Agent sessions and Archivum ingest call
a model provider unless you configure a local one.
Why it is built this way
Coding agents and computer-use agents want broad access — your whole repo, a real
desktop, your notes. That is a reasonable thing to grant a tool running on your
own hardware and an unreasonable thing to grant a SaaS product. Keeping the
control plane local is what makes the access level acceptable.
It also means the failure modes are yours to fix. Nothing is gated behind a status
page you cannot see. 
