Tarralappu (Finnish: sticky note / label) is a sticky-note-themed harness for AI image generation via OpenRouter. Bring your own API key, pick a model, write a style guide, attach reference images, and generate variations in a left-to-right chain of refinements.
Live at lumikeiju.dev/tarralappu/.
- Model-aware: dynamically fetches every image-capable model from OpenRouter, grouped by creator with inline pricing and an at-a-glance cost tier.
- Unified generation: uses OpenRouter's dedicated Image API for discovered models and keeps chat completions as a compatibility path for legacy models.
- Chain-based workflow: each chain (row) locks to one model. Add a root sketch, refine it to the right, fork any card into a new row to branch.
- Prompt Board: a scratchpad of reusable prompt notes; "send" any note to one or more models at once, each landing as an editable draft.
- Global style context: a style document, style reference image, and layout reference image attach to any sketch via checkboxes.
- Conversational refinement: reasoning models receive the full thread history; image-only models (Flux, etc.) automatically fall back to image-to-image.
- Image controls: model-aware references, output count, resolution, quality, format, compression, seed, and optional SSE previews where supported.
- Cost tracking: estimated and actual cost per sketch, chain totals, session total, configurable caps.
- Clear error messages: typed API errors are mapped to plain language, with a raw-response view for debugging.
- Client-only, no backend: your API key stays in the browser and is sent directly to OpenRouter.
- Dark and light sticky-note themes, WCAG 2.2 AA compliant.
bun install
bun run devThen open http://localhost:5173/tarralappu/.
bun run build # type-checks then bundles to dist/
bun run preview # serve the production build locallyPushing to main triggers the GitHub Actions workflow, which builds with Bun and publishes dist/ to GitHub Pages at lumikeiju.dev/tarralappu/.
Full usage documentation lives in Silmärin.
All repository conventions live in AGENTS.md, which both humans and AI coding agents should read first. The .github/ files implement those conventions for Copilot, Codex, and similar tools.