The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
/**
* Matching an image/video model NAME typed in a chat call against the models
* actually installed for the macOS MLX lane, plus the rule for what an
* argument-less call falls back to.
*
* It lives here rather than in api/mcp/builtin-tools.ts because that module
* closes an import cycle through the tool registry and cannot be loaded by a
* test (see lib/__tests__/tool-classification.test.ts, which reads it as
* text). Pure functions, no store and no I/O, so both rules are testable.
*/
/** The shape both MLX catalogs (image and video) share. */
export interface NamedModel {
id: string
name: string
}
/**
* Fuzzy-resolve a chat-typed model name against an installed MLX catalog by id
* or display name. Same tolerant-matching shape as resolveModelName in
* api/vram-handoff.ts, kept separate so this Mac-only path does not pull in
* the ComfyUI-flavoured module.
*/
export function resolveMlxModel<T extends NamedModel>(
requested: string | undefined,
installed: T[],
): T | null {
if (installed.length === 0 || !requested) return null
const norm = (s: string) => s.toLowerCase().replace(/[^a-z0-9]+/g, '')
const r = norm(requested)
if (!r) return null
return (
installed.find((m) => norm(m.id) === r || norm(m.name) === r) ??
installed.find((m) => norm(m.id).includes(r) || norm(m.name).includes(r)) ??
installed.find((m) => r.includes(norm(m.id)) || r.includes(norm(m.name))) ??
null
)
}
/**
* Which installed MLX image model an argument-less `image_generate` runs on.
*
* Same rule the ComfyUI gate follows in api/model-pick.ts, kept in step for
* cross-platform parity: the Create tab holds the model the user picked and
* can see, so it answers first. Nebenbefund N1 of the D1 counter-check
* (Windows build, 2026-08-29) caught the ComfyUI half of this ignoring that
* choice; the MLX half had the same shape.
*
* sd-turbo stays the fallback for an install that never opened Create, then
* the first installed model. Undefined only for an empty install, which the
* caller reports with its own message.
*/
export function defaultMlxImageModel<T extends NamedModel>(
installed: T[],
createChoice: string | null | undefined,
): T | undefined {
return (
resolveMlxModel(createChoice || undefined, installed)
?? installed.find((m) => m.id === 'sd-turbo')
?? installed[0]
)
}