Architecture and trust boundaries
Native desktop shell (Electron, isolated renderer)
└─ owned loopback backend (Python)
├─ project grants / approvals / goals
├─ knowledge extraction → compact local embedding/index
├─ agent orchestration → configured API or local model endpoint
├─ tools → scoped files / approved execution / optional connectors
└─ artifacts + conversation state + provenance ledger
Phone clients → authenticated messaging gateway / private dashboard path
The desktop owns backend startup and cleanup. Packaged runtime paths are separate from per-user profile paths. It does not require the judge to open a web page or start two servers manually. The renderer uses isolated context, disabled Node integration and a constrained native folder-picker bridge.
The model is not a trusted operating-system principal. Tool calls still pass through configured approvals and scopes. Retrieved content is data, not policy. Provider calls cross a network trust boundary; private phone sharing has its own authentication boundary. Optional local inference moves model execution to a local server but does not eliminate tool risk.
Main code areas: origo-desktop, origo_runtime, and the imported agent/workbench under indra/SIH-26/hermes-agent. Internal compatibility identifiers are retained intentionally. See development for running and testing the source.