Most people meet a coding agent as a product they log into. The product owns the loop, the tools, the transcript, the permission prompt, and the account. That is a good way to get work done and a bad way to learn what a harness is. Firebrand is the other move. It is a terminal program that runs prompt, model step, tool calls, tool results, repeat — against a workspace — with its own tools, its own subagents, and an append-only session record. A model is a leg: an endpoint plus an auth source plus a wire dialect. The loop never knows which. The primary planes are a local model through a local daemon and a cloud model through that same daemon after the operator has signed in there. Other brains are compatibility surfaces, not the point. Bare is the default. Identity, when attached, is a three-mode dial — bare, mini, full — resolved by model class, overridden by a workspace door file, and refused rather than trimmed when the closure is broken. The record is the lab notebook: everything the model could see, written before it is shown, folded back into a transcript on resume. The Mac app supervises the same engine over a stdio bridge and opens no engine port. The mobile app is the family's first native flight onto a phone, reached through one opt-in remote gateway. Anvil is the measuring instrument; Firebrand is a specimen that can join a run. Kiln is a workshop that never authors; this loop does author, and that difference is the whole contrast. Conceptual open-sourcing: the architecture is the deliverable, not a repository you clone.