Report which of the three execution modes this server can run RIGHT NOW given the current env: inline_svg (zero key — hosting LLM authors the SVG), external_prompt_only (zero key — paste prompt into Ideogram/Nano Banana/Midjourney/Recraft/Flux UIs, then asset_ingest_external), api (requires provider key). Read-only; no network. Call before offering the user options.

Classify an asset brief, route to the right model, rewrite the prompt in that model's dialect, and report which execution modes are available (inline_svg / external_prompt_only / api). Returns an AssetSpec JSON including modes_available, optional svg_brief (for inline_svg), optional paste_targets (for external_prompt_only), and — when the brief leaves a material ambiguity — a clarifying_questions[] array the host LLM should surface via AskUserQuestion (or the equivalent) BEFORE calling a generator. Each entry has {id, header, question, options[], required, why}. Read-only; idempotent; no network.

Generate a logo. Three modes: inline_svg (Claude emits SVG — zero key), external_prompt_only (returns prompt + paste targets — zero key), api (server runs the provider pipeline — requires key). Omit mode to auto-select. Returns an AssetBundle / InlineSvgPlan / ExternalPromptPlan discriminated by the mode field.

Generate an app icon (iOS AppIconSet, Android adaptive, PWA maskable, visionOS 1024² master + placeholder parallax layers). Three modes (inline_svg / external_prompt_only / api). In non-api modes only the master mark is produced; call asset_ingest_external afterwards to run the platform fan-out. Set ios_18_appearances=true to also emit dark and tinted 1024² variants for iOS 18 tintable icons.

Generate a favicon bundle (favicon-{16,32,48}.png + icon.svg + icon-dark.svg + apple-touch + PWA 192/512/512-maskable + snippet). Three modes — inline_svg is the best fit for simple glyph marks (legible at 16×16).

Render a 1200×630 OG image via Satori template (deterministic typography, no diffusion). Default mode=api renders server-side without any API key. external_prompt_only is only meaningful when with_background_image is set. inline_svg is not supported (web-font loading + precise text layout beyond LLM reach).

Generate one or more brand-locked illustrations. Two modes (external_prompt_only / api); inline_svg is not supported — path budget too small for a composed scene. Injects brand bundle (palette, style_refs, LoRA, style_id) where supported.

Generate a cross-platform splash-screen bundle from a brand mark. Two modes (external_prompt_only / api); inline_svg is not supported (splash screens are PNG bundles — generate a logo inline_svg first, then call this with existing_mark_svg). api mode composites the mark onto background_color and emits ios/LaunchScreen-2732.png, android/mipmap-*dpi/splash.png, android/themes-splash.xml, pwa/splash-1200.png, and a README describing how to wire each.

Generate marketing-hero / landing-page banner art. Two modes (external_prompt_only / api); inline_svg is not supported. Accepts aspect_ratio (16:9 / 21:9 / 3:2 / 2:1). Injects brand bundle where supported. Returns N variants if count>1.

Matte an image to transparent background (BiRefNet / BRIA RMBG / U²-Net via remote endpoint; local white-chroma fallback). Returns RGBA PNG path.

Convert a raster image to SVG. Tries in order: Recraft /vectorize (if PROMPT_TO_BUNDLE_RECRAFT_VECTORIZE_URL is set), vtracer on PATH, potrace on PATH, then a built-in posterize run-length fallback. Passes all output through SVGO when installed.

Upscale / refine an image, asset-type-aware. DAT2 for flat logos/icons, Real-ESRGAN/SUPIR for photoreal, img2img for diffusion polish. Lanczos fallback.

Run tier-0 deterministic validators on an asset (dimensions, alpha presence, checkerboard-pattern heuristic on tile-luma alternation, safe-zone bbox, palette ΔE2000 against brand, WCAG contrast of brand primary vs light and dark tabs, OCR Levenshtein against intended_text). Optional tier-2 VLM-as-judge via PROMPT_TO_BUNDLE_VLM_URL.

Parse a brand source (brand.json, DTCG tokens, AdCP spec, brand.md, or raw text) into the canonical BrandBundle schema.

Round-trip endpoint for inline_svg mode. After you (the LLM) emit the … in chat, immediately call this tool with that SVG text so the server writes a complete asset bundle to disk: master.svg + (for favicon) icon.svg + icon-dark.svg + favicon-{16,32,48}.png + favicon.ico + apple-touch-icon.png (opaque) + pwa-192.png + pwa-512.png + pwa-512-maskable.png + manifest.webmanifest + head-snippet.html + (for app_icon) the full iOS AppIconSet + Android adaptive (foreground+background+monochrome) + PWA maskable + visionOS. Returns an AssetBundle with file paths the user can open. Validates the SVG against the original svg_brief (viewBox, path count, palette, forbidden elements).

Ingest an image the user generated in an external tool (Midjourney, Nano Banana, Ideogram web, Recraft, Flux Playground, etc.) and run the matte → vectorize (where applicable) → tier-0 validation pipeline. The round-trip endpoint for external_prompt_only mode.

Train a brand-consistent LoRA from 20-50 sample images, returning a lora_id the comfyui-* and SDXL-family providers can reference. Requires a user-owned training endpoint (Modal / Runpod / self-host) at PROMPT_TO_BUNDLE_MODAL_LORA_TRAIN_URL. Phase-4 scaffold: the MCP tool does the packaging, validation, and HTTP; the user owns the deployment and pricing. See docs/research/06-stable-diffusion-flux/6d-lora-training-for-brand-style.md.

Structured environment inventory — MCP equivalent of p2a doctor. Returns native-dependency status (sharp, vtracer, potrace, png-to-ico, satori, resvg-js, tesseract.js, svgo), free-tier routes ranked best-first, paid-provider keys, paste-only providers, pipeline extension URLs, which modes are available right now, and a concrete 'what to try next' suggestion list. Read-only by default. Pass check_data=true to also run the model-registry/routing-table integrity check. Pass auto_fix=true to install missing native binaries (Homebrew / cargo / scoop — never sudo); pair with auto_fix_dry_run=true to preview without executing.

List the model registry (60+ entries) with optional filters. MCP equivalent of p2a models list. Returns id, family, provider, dialect, native_rgba/svg flags, text ceiling, tier (free/paid/paste-only), key_set status. Filter flags: free, paid, paste_only, rgba, svg. Read-only; no network.

Full capability dump for one model. MCP equivalent of p2a models inspect <id>. Accepts a model id or an aka alias. Returns the full ModelInfo record, env status, paste targets, routing rules that reference this model (as PRIMARY / fallback / NEVER), and usage notes. Read-only; no network.

Fan out a 1024² master PNG into the full platform bundle (iOS AppIconSet, Android adaptive, PWA maskable, visionOS parallax, Flutter launcher, favicon set). MCP equivalent of p2a export master.png. No API key required; runs entirely on sharp. Use when the LLM has a master (inline_svg saved, api-mode result, or user-supplied hand-authored PNG) and needs the platform fan-out.

Pack a directory of PNG/WEBP/JPG frames into one sprite sheet + TexturePacker-compatible JSON atlas (works in Phaser, PixiJS, Three.js, Godot, Unity via a light importer). MCP equivalent of p2a sprite-sheet <dir>. Offline, no API key.

Emit a 9-slice config + CSS border-image snippet + Unity/Godot/Phaser/PixiJS-ready numbers from one image and 4 pixel offsets. Optionally also emit an Android .9.png with the 1px stretchable-region encoding. MCP equivalent of p2a nine-slice <image>.

Scaffold brand.json in the project root + ensure the assets dir exists. MCP equivalent of the brand.json portion of p2a init. Auto-detects the framework (Next.js, Expo, Flutter, Xcode, Astro, Vite, Remix, Nuxt, React Native, Electron, Node) and returns platform hints. Deliberately does NOT do IDE MCP registration — that's the one piece the user handles once at install time via a terminal. Call this at the start of a new project so subsequent generator calls have a brand source-of-truth and a known output dir.