The problem was consistency, not another chat box

Kora can be used from a Tauri desktop application, a command-line interface, and a Chrome extension. Each surface has a different interaction model, but the underlying work—retrieving context, selecting tools, updating state, and resuming a session—must stay consistent.

Putting a separate agent loop inside every interface would have made each surface easy to start and difficult to trust. Tool behavior, prompts, memory rules, and session state would drift as soon as one client changed faster than the others.

The boundary I chose

The interfaces request work and render events. The runtime owns the behavior that must remain coherent: session control, tool and skill execution, context retrieval, and durable state. That also keeps private context on the user’s machine by default.

What the design improves

  • Continuity: a session can be resumed without teaching each interface its own memory model.
  • Inspectability: one owner records tool activity, retrieved context, and durable state.
  • Consistency: tools and skills behave the same way regardless of where a request begins.
  • Evaluation: retrieval and safety behavior can be tested below the interface layer with repeatable fixtures.

What the design makes harder

A shared runtime creates a real contract between clients and the system underneath them. Events need stable shapes. The browser extension cannot depend on desktop-only internals. Packaging, local process management, recovery, and version compatibility become product problems rather than details hidden inside one application.

Local-first storage also improves user control while making cross-device synchronization more difficult. I chose that tradeoff deliberately: Kora’s default responsibility is to the user’s machine and context, not to a hosted account.

How I check the boundary

Kora’s current retrieval evaluation uses 185 synthetic search scenarios. The measured result is 99.2% ranking quality, with every correct result appearing in the top five and no safety or data-integrity failures in that suite. That evidence supports the retrieval boundary; it does not claim universal model accuracy or complete production reliability.