A local-first Windows assistant that turns natural-language requests into multi-step actions across personal knowledge, schedules, projects, documents, and browser or desktop applications.
Designed and built by Mobolaji Ogunbiyi · December 2025—Present
A sanitized fixture demonstrates the implemented control flow without connecting to private accounts or a live model. Use the controls to inspect each step; the complete reading path remains below without JavaScript.
Kora resolves the requested outcome into a bounded sequence: identify the application, retrieve candidate evidence, research the company, assemble the brief, and save the artifact.
goal: interview_brief steps: resolve → retrieve → research → compose → save write_scope: applications/interview-brief.md
The runtime searches durable local sources and returns supporting spans rather than relying on conversation memory alone.
sources: master resume, application record, project evidence result: 7 cited spans selected
The agent reads the relevant browser page, records the company context, and prepares a document through typed tools. Effects remain explicit and attached to the active run.
browser.read_page → engineering principles document.compose → structured interview brief filesystem.write → applications/acme/interview-brief.md
The final response reports what changed, where the artifact lives, and which sources support it. The session retains enough state to continue after compaction or restart without replaying effects.
completed: true artifact: applications/acme/interview-brief.md citations: 7 next: review system-design talking points
These results describe the current automated retrieval evaluation—not universal model accuracy or production reliability.
The suite uses synthetic personal histories. Every correct result appeared within the top five. Separate live tool-selection and operational qualification remain different evidence boundaries.
One authenticated runtime shared by a Tauri desktop app, CLI, and Chrome extension so the assistant can keep context across surfaces.
Kora-native orchestration, typed tools and skills, context retrieval, durable SQLite state, and resumable sessions around the Pi runtime.
Synthetic histories, repeatable queries, ranking checks, safety assertions, and data-integrity assertions used to guide implementation.
Product surfaces share one runtime and durable state instead of duplicating agent behavior in each interface.
Personal context and durable state stay on the user’s machine. That improves control and inspectability, but makes packaging and cross-device synchronization harder.
Sharing a runtime avoids behavioral drift. It also requires strict surface contracts so desktop, browser, and CLI clients do not leak internal fields.
Retrieval is measured with repeatable fixtures. A strong offline result does not substitute for connected, long-running, or production evidence.
Kora is an active private project with a working local runtime, desktop and browser surfaces, durable sessions, typed tools, retrieval, and automated evaluations. It is not presented here as a public hosted service or open-source repository.