Sovereign and offline AI
Sovereign AI means running AI on infrastructure you control, under your own jurisdiction and rules, without depending on a service you do not control. The knowledge an AI answers from has to work the same way: it cannot rely on an outside search service, a hosted index, or a call home to check where it came from.
MOCA grounding runs entirely where you put it. A MOCA package is a file. A MOCA Reader is a library that runs inside your application. Reading, verifying, searching and citing need no network, no external service and no particular model.
What runs where
Section titled “What runs where”| Step | Where it runs | Needs |
|---|---|---|
| Read a package | In your process | The .moca file or folder |
| Verify its digest and signatures | In your process | A trust root you control, listing the keys you accept |
| Check quotes against their sources | In your process | Nothing: the originals travel inside the package |
| Lexical search | In your process | Nothing |
| Semantic search | In your process | A sidecar index, and an embedding model you run yourself |
| Answer a question | Your model, wherever you run it | Your choice of model |
The rules that make it hold
Section titled “The rules that make it hold”These are part of the Reader contract, which every Reader must follow, and the conformance corpus tests:
- A Reader never reaches the network because a package names something. A package can cite a web address, but the Reader never fetches it.
- A package never configures its Reader. No field can make a Reader load a model, call a service or change a setting.
- The host decides whom to trust. Signatures count only against a trust root you supply. DSSE signatures with Ed25519 keys verify offline in every Reader.
- Loading a package never runs anything. A package is data.
Air-gapped, edge and on-device
Section titled “Air-gapped, edge and on-device”- Air-gapped networks. Carry a signed
.mocafile and its sidecar index across the gap; verify it on the other side against your trust root. - Edge and devices. Readers are libraries in TypeScript (Node.js, Workers, browsers), .NET and Python. A package with a lexical index searches on a laptop or tablet with nothing else installed.
- Your own models. Semantic search uses an embedding model you choose and run, such as a small local model through Ollama or ONNX Runtime. The index records the model it was built with, and the Reader refuses a mismatch (search and sidecar indexes).
- Your own agents. The MCP server runs as a local process on the same machine as the client.
Proving what was used
Section titled “Proving what was used”Sovereignty also means accountability. With MOCA you can show exactly which knowledge an assistant used, long after the fact: log the package digest and the versioned page reference with each answer. Together they name the exact text, independent of any index or service that may since have changed (use cases: audit and regulated AI).
- Architecture: the package, the Reader, and what the host decides.
- Signing and review: keys you control.
- Consuming packages: read once, verify once, search and cite.