Air-gapped by design: AI caching where the internet isn't invited

Air-gapped and egress-restricted environments run a brutal vendor filter: cloud control planes, online license checks, telemetry beacons, and package-manager installs all die at the firewall. Most modern AI tooling fails the filter in the first paragraph of its own docs, the architecture assumes a chatty relationship with the vendor.

Crowkis assumes silence. The binary is self-contained; the license verifies with a local Ed25519 signature check against a baked-in public key, our servers are not consulted, ever; there is no telemetry to disable because none exists; and the dashboard, metrics, and full intelligence stack run entirely within the gap. Install media is one image file you carry in.

flowchart LR
  subgraph IMG["crowkis/crowkis:latest"]
    B["one stripped Rust binary"]
    D["/data volume"]
    U["non-root user"]
  end
  subgraph NOT["deliberately absent"]
    N1["no Python · no PyPI"]
    N2["no package manager"]
    N3["no dependency tree"]
  end
  IMG ~~~ NOT
  style IMG fill:#fbe9e8,stroke:#d62221,stroke-width:2.5px

The workload fit is stronger inside the gap, not weaker: constrained or accredited inference capacity makes every avoided model call precious, and the fixed-corpus analyst workloads typical of these environments are repetition engines. Sub-millisecond hits stretch scarce compute exactly where procurement cycles can't.

#The bottom line

Most vendors retrofit an 'offline mode' as a sales checkbox. Offline is our only mode, the connected deployment is just the air-gapped one with friendlier neighbors.