SmoothyStudio

Status: macOS only for now. Windows and Linux targets are already configured in package.json, but nothing is built or tested against them yet.

A focused thumbnail tool: generate the imagery, compose the thumbnail, export it — plus an MCP server so an agent can do the whole loop for you.

Two surfaces in one window:

• Studio — the canvas designer. Text (single / bullets / batch / split), shapes, icons, pictures, borders, corner logos, and a reorderable per-layer effect pipeline (crop, mosaic, wave warp, mirror, colour replace, duotone, roughen, blurs, vignette, turbulent displace, echo, venetian blinds, transform, 3D tilt, and a Lumetri-style grade). Exports transparent PNGs, one at a time or a whole batch.

• AI Photos — generation for the assets a thumbnail needs: reaction faces, text-to-image, background removal, upscaling, vertical covers.

The two are wired together: Send to Studio on any generated image drops it in as a picture layer, and the picture layer has a Generate with AI button that takes you the other way.

Running it

Requires Node 20+ and macOS.

npm install
npm run dev

Signing in uses your existing SmoothyEdit account — the app opens your browser and receives the token back, so no credentials are stored in this repo or entered into the app.

Building a release

npm run dist:mac   # → dist/SmoothyStudio-<version>.dmg and .zip

Builds are unsigned for now, so the first launch needs one confirmation: on macOS 15+ open it, click Done, then System Settings → Privacy & Security → Open Anyway; on macOS 13–14 right-click → Open.

Dev server runs on 127.0.0.1:5176, so it can run alongside SmoothyDesktop (5175).

npm run typecheck   # node + web
npm test            # vitest
npm run build

Business model

The editor is free. Image generation costs credits, metered against the same /api/credits rail the web app already uses — inference has real marginal cost, composition doesn't.

Background removal runs on device and costs nothing. It uses the ISNet model bundled with @imgly/background-removal-node in a utilityProcess, so photos never leave the machine, there is no per-image cost, and there is no upload size limit — a full-resolution PNG exceeds a hosted function's request-body cap before it even starts.

MCP server

The app hosts an MCP server on http://127.0.0.1:3777/mcp while it's running. Rendering happens in the app's own renderer (through window.__studioMcp), so agents get the exact code path the UI uses — no second renderer to keep in sync.

Connect Claude Code to it:

claude mcp add --transport http smoothystudio http://127.0.0.1:3777/mcp

Then ask for thumbnails in plain language. Three tools, deliberately coarse — a tool per effect would make an agent burn turns and produce incoherent layouts:

Agent renders land in the app as new canvases; nothing you already have open is touched or overwritten.

The document model

There is no parallel "spec" format to learn. A Studio document is plain JSON, and every tool takes a partial document that's merged one level deep over the defaults — or over a template when you pass templateId:

{
  "templateId": "bold-centered",
  "doc": {
    "text": "MY HOOK\nGOES HERE",
    "font": { "family": "SF Pro Display", "weight": 700, "size": 180, "color": "#FFFFFF" }
  }
}

Templates exist because an agent handed a blank canvas and raw coordinates produces valid but ugly layouts. Each one fixes the composition — type scale, safe margins, contrast — leaving the words and the imagery to the model. Current set: bold-centered, subject-right-text-left, top-banner, vertical-short.

Rendered PNGs come back inline so a vision-capable model can look at its own output and iterate. That feedback loop is the point; a tool that only returned a file path couldn't close it.

Placement. Don't guess coordinates — measure returns real boxes, and anchors compute positions for you:

{ "anchors": [{ "id": "sub", "width": 520, "height": 54,
                "anchor": { "to": "text", "edge": "below", "gap": 40 } }] }

Masking. Any layer can be clipped to another layer's silhouette — imagery through knocked-out type, a photo inside a shape:

{ "image": { "fx": { "mask": { "enabled": true, "sourceId": "text" } } } }

The finishing pass. doc.canvasFx and doc.canvasGrade apply to the composite, not one layer — a vignette on doc.fx darkens the title's letters, which is almost never what you want.

Effects live on a layer's fx object and the colour grade on its grade object, both optional:

{ "doc": { "text": "BOLD CLAIM",
           "grade": { "enabled": true, "contrast": 130, "vibrance": 25 },
           "fx": { "vignette": { "enabled": true, "amount": -45 } } } }

Where your work is stored

Canvases are written to studio-workspace.json in the app's userData directory, not localStorage. Documents embed their images as data URLs, so a couple of generated pictures blow past localStorage's ~5 MB quota — and it fails silently, losing work. The file-backed path has no cap, writes atomically, and surfaces any failure in the Studio header.

Layout

src/main/        Electron main — auth, credits, AI Photos bridge, export dialogs, MCP server
src/preload/     contextBridge surface (auth, credits, aiPhotos, studioApi)
src/renderer/    React UI — Studio + AI Photos, the render library, the MCP bridge
src/shared/      Types shared across processes

The rendering core is src/renderer/src/lib/studio.ts (document model + pure renderer) and studio-effects.ts (the effect pipeline). renderDocToPngBase64(doc) is a pure document-to-pixels function, which is what makes the whole thing drivable by an agent.