# One runtime for the whole desk (the thesis)

The typical AI desk = five tools taped together (chat subscription, terminal assistant, local runner, scheduler) with no shared memory, tooling, or audit trail. Kaptain's thesis: that belongs in ONE local runtime.

## Why taped-together fails
Separate context, permissions, bills; nothing composes — the scheduler can't use the strong model's judgment, the terminal assistant can't see the project map, the local runner idles while the subscription burns.

## One runtime
A single binary: workbench + every model route + toolbelt + agents/schedules + observability, on your machine, inside a chosen project boundary. Everything composes: a scheduled kapsule can delegate to a local model over the mesh with the same tools and audit trail as live chat.

## Model-agnostic, concretely
Four routes in one picker (/knowledge/choosing-model-routes/) + other machines' models (/knowledge/model-lending/), swappable per chat / kapsule / fleet.

## Receipts
https://chynj.ca/built-with-kaptain/ — real apps built end to end inside this runtime.