Pixel-art asset generation for game projects, MCP-first via the connected pixellab MCP server (mcp__pixellab__* tools). Triggers on: "generate pixel art", "make a pixel sprite/tile/character", "animate this sprite", or any pixel-art asset request for app/games/**. Falls back to scripts/pixellab.mjs only for animated WebP assembly or as a REST emergency path.
Pixel-art asset generation for game projects, MCP-first via the connected pixellab MCP server (mcp__pixellab__* tools). Triggers on: "generate pixel art", "make a pixel sprite/tile/character", "animate this sprite", or any pixel-art asset request for app/games/**. Falls back to scripts/pixellab.mjs only for animated WebP assembly or as a REST emergency path.
PixelLab Skills (MCP-first)
Pixel-art asset generation for app/games/**. The PixelLab MCP server is the default
path — it is connected this session as mcp__pixellab__* tools, lands every asset in the
PixelLab dashboard (reusable, style-chainable), and bills against our subscription. Use the
local scripts/pixellab.mjs only for the one thing the MCP can't do (assemble an animated
WebP) and as an emergency REST fallback.
MCP-first. Generate through mcp__pixellab__*. Reach for scripts/pixellab.mjs only for
webp assembly or if the MCP server is unreachable.
Write BIG, exact descriptions. PixelLab rewards long, fully-specified prompts: subject,
view angle, palette, lighting, material, silhouette, and an explicit list of what to EXCLUDE.
Vague prompts waste generations. (See "Writing prompts".)
Style-lock related assets with create_object_state / create_character_state on a
completed parent — never cold-generate siblings (loses palette/identity). See [[feedback-pixellab-card-state-parent]], [[feedback-claimer-sprite-refs]].
Never save binary with the Write tool — it corrupts PNGs to zero-filled garbage.
curl -L the rotation URL, then file-verify.
Bump the filename on any asset replacement (-v3 → -v4) — public/assets/** is
immutable-cached 1y + 30d nginx. See [[feedback-nginx-static-cache]].
PixelLab only — no PixelForge fallback for game sprites. See [[feedback-pixellab-only-for-games]].
⚠️ Tool-name reality check
There is NOcreate_object and NOvary_object tool — those names (in older memory) were
never real. The actual primitives on the connected server are:
Generation is asynchronous. The create call returns an id immediately; you poll until
completed, then download the PNG from a public URL.
1. create_1_direction_object({ description, size }) → returns object_id
2. get_object({ object_id, include_preview: false }) # poll every ~10–30s
status: processing → progress + ETA, keep polling
status: review → candidate grid (see "Review gate") — select or dismiss
status: completed → has `rotations:` + `download:`
status: failed → error + retry hint
3. Download the PNG. `get_object` (completed) returns, exactly:
rotations:
<dir>: https://backblaze.pixellab.ai/file/pixellab-characters/objects/<grp>/<id>/rotations/<dir>.png
download: https://api.pixellab.ai/mcp/objects/<id>/download # zip archive of all frames
For a 1-direction object the rotation key is literally `unknown`.
4. curl -L "<rotation-url>" -o public/assets/games/<game>/<name>-v1.png # NO auth header needed
5. file public/assets/games/<game>/<name>-v1.png # MUST say "PNG image data, WxH, 8-bit/color RGBA"
Download facts (verified live): rotation URLs are public Backblaze links — a plain
curl -L with no Authorization header returns the real PNG (HTTP 200). The MCP never
returns raw bytes and never writes files; it only hands back URLs (+ an optional inline
preview when include_preview is true — for eyeballing, not for saving). Poll budgets:
1-dir/tiles ~15–90s, 8-dir/characters ~2–4 min, tilesets ~100s.
Review-status gate (don't lose your asset)
create_1_direction_object with size ≤ 170 returns multiple candidates in review
status (≤42→64 candidates, ≤85→16, ≤170→4). get_object shows them inline; you must then
call select_object_frames({ object_id, indices: [n] }) to promote the good one(s) into their
own completed objects, or dismiss_review to discard. Leave it and the asset never finalizes.
Use size ≥ 171 to get a single candidate with no review step.
Style consistency: object-states & style-refs
Build a family of related sprites by editing a completed parent, not by re-prompting:
create_1_direction_object({ description: "<the first/base sprite>", size: 48 }) → parentId
create_object_state({ object_id: parentId, edit_description: "make it a green shell instead" })
create_object_state({ object_id: parentId, edit_description: "make it a lightning bolt instead" })
All variants share palette/outline/identity via an automatic group_id. This is the
proven pattern for the mau-mau card faces (parent 06f843eb-abc3-4b2e-a7ee-c2974acbebb1 —
create_object_state ONLY, never a fresh create_*) and the claimer directional sprites.
For style-matching across a totally new object, pass style_images to create_1_direction_object
(largest style image's size sets output size — don't also pass size).
Writing prompts (the single biggest quality lever)
PixelLab interprets a long, exact description far better than a terse one. Always specify:
subject + view + palette + lighting/shading + material/texture + silhouette + EXCLUSIONS.
✅ Good: "Single top-down pixel-art banana-peel power-up icon, 48×48, lineless flat shading,
transparent background. Glossy yellow curved peel lying flat as if dropped on a track, small
brown bruise tip, soft inner neon-yellow rim glow, dark drop-shadow blob beneath. Centered,
no text, no border, no frame."
❌ Bad: "a banana item" — model guesses size, view, palette, adds a background/border.
outline: "lineless" almost always (user hates borders).
For opaque scenes add "edge to edge, full bleed, no border, no frame, no padding" or you get a
white-canvas vignette. For "no characters" say so explicitly — models populate scenes by default.
detail: "medium detail", shading: "medium shading" are good defaults. Use seed for reproducibility.
Cost model (dual bucket — the old "16 credits/call" rule is WRONG)
get_balance shows two separate buckets:
Subscription generations — the primary bucket (Tier 1 = 2000 per billing period).
Each call decrements this first.
USD credits — a fallback only, used when generations are exhausted.
Cost is per call, by endpoint + size — NOT per image. One v2 call can return a grid of up to
64 images for the same generation cost. So:
Standard object/character = 1 generation. v3 character = 2–9.
create_1_direction_object / create_8_direction_object use Pro Tools = ~20–40 generations per call by size.
The only real cost cliff is Pro-mode animation: 20–40 generations PER DIRECTION (a full
8-dir Pro animation = 160–320 = ~6–12 of those empties the monthly allowance). Prefer
mode: "v3" (the cheap default, often higher quality). For Pro animate_*, call withoutconfirm_cost first to get the quote, surface it, then re-call with confirm_cost: true.
enhance_prompt silently adds 0.05 generations.
Check get_balance before a big batch; budget by calls, not by samples.
File-saving rules
Never use the Write tool for image data (corrupts binary → zero-filled file).
Download via curl -L "<rotations.<dir>>" -o <path> (no auth header). The download: URL is a
zip archive of all frames — use it only when you want every rotation/frame at once.
Alwaysfile <path> — must report PNG image data, …, RGBA. If it says data/JPEG/RIFF,
it's broken; regenerate, don't patch in place.
Optimize after: pngquant --force --quality=60-85 --ext .png <path> (60–85% is indistinguishable
for pixel art; a 256×384 bg drops ~130KB → ~50KB).
Replacing an existing asset? Bump the filename suffix and update the loader key (nginx cache).
Animation & animated thumbnails
In-engine motion:animate_object({ object_id, animation_description, mode: "v3" }) (1 gen/dir,
frame_count 4–16, default 8 → stores 9 incl. a ref frame). Poll get_object; completed response
lists animations: with downloadable frame URLs (Backblaze, same as rotations).
Animated thumbnails / menu loops: download the frames, then assemble locally with
scripts/pixellab.mjs webp --in-dir <frames> --out <x>.webp --duration 120. The MCP/REST cannot
assemble a WebP — this is the one irreplaceable local step. See [[feedback-pixellab-animate-webp]].
Pixel-exact thumbnails (compose the game's REAL sprites with overlay sparkle/pulse, then WebP):
prefer this when identity must be preserved — generative animate-with-text-v3 reinterprets the
concept and will NOT keep your exact pixels. (The reason is pixel-identity, not cost.) Full PIL recipe in
[[feedback-pixellab-animate-webp]] / the api-v2-extra reference.
The audit confirmed its commands are valid REST calls; it is superseded for static work, not
broken. Keep it. Reads PIXELLAB_API_KEY from .env (already set).
Command
Status
webp
LOAD-BEARING — local img2webp (libwebp); no MCP/REST equivalent.
Hard host prerequisite for webp: img2webp from libwebp (brew install webp). One genuine use
where the script wins: a wide panoramic banner (e.g. a skybox strip) — MCP object tools pad
toward square, so pixellab.mjs sprite with an explicit wide --size is the right tool there.
⚠️ But generate-image-v2 caps aspect at ~688×384 (16:9) — extreme strips (e.g. 1024×192) exceed it;
generate a seamless tile within the cap and let the engine TileSprite repeat it.
Tilesets — Wang (top-down) PREFERRED for terrain
create_topdown_tileset → 16 Wang tiles (23 if transition_size: 1.0), 32×32, outline: "lineless".
lower_description = ground, upper_description = elevated, transition_description = edge blend.
~100s. Chain seamless multi-terrain via the returned base tile IDs.
Download:curl -L --fail (302 redirect → 0 bytes without -L).
Wang index + Phaser integration (vertex-terrain algorithm, frame lookup, RenderTexture layering)
— unchanged and correct; see the detailed recipe in the api-v2-extra reference and below.
Wang index = NW×8 + NE×4 + SW×2 + SE×1 (upper=1, lower=0)
Frame = (bbox.y/32)*4 + (bbox.x/32)
Vertex(vr,vc) = 0 if ANY surrounding cell is target terrain, else 1
Tiles Pro / isometric:create_tiles_pro (number each variant: "1). grass 2). dirt …",
n must match) / create_isometric_tile (tile_shape thin/thick/block). Small pickups look
BAD as tiles — use Graphics primitives for tiny in-game objects; reserve tiles for 32×32+ decor.
Corner-tile orientation is often mislabeled — verify at 10× zoom, or just use Wang.
Characters & enemies
Humanoid / quadruped:create_character (body_type, mode: standard|v3|pro, n_directions 4|8).
Standard = 1 gen; v3 = 2–9 (always 8-dir, top quality); pro = 20–40. Canvas is ~40% larger than the
requested size (leaves animation room) — slice sheets accordingly.
Variants (armor color, sitting, damaged): create_character_state (consistent across all rotations).
Animate:animate_character — template_animation_id (walk/run/idle… 1 gen/dir) or
action_description (custom, v3). South = default facing; West = East with setFlipX(true).
Blobs / slimes / non-humanoid: NOT create_character (forces legs). Use create_1_direction_object
/ create_map_object per pose, describe "no legs, no arms, blob body".
Gotchas
Base64 image inputs use { "type": "base64", "base64": "<png-bytes-b64>", "format": "png" } —
base64-encoded PNG, not raw RGBA, not a data URL.
Size limits vary per endpoint: generic min 16×16; animation/rotation max 256×256;
generate-image-v2/style/ui square ≤512 (16:9 ≤688×384); animate-with-text v1 locked to 64×64;
tiles-pro 16–128. Let the API's 400 error tell you the real max for your aspect ratio.
One v2 image call returns a GRID (up to 64 tiny images), not a single file — parse accordingly.
no_background defaults FALSE on legacy create-image-*, TRUE on the v2 Pro endpoints.
Raw-RGBA vs PNG: some REST responses are base64 raw-RGBA, not PNG; pixellab.mjs auto-wraps
these. MCP rotation URLs are always real PNGs (verified). Always file-check.
Destructive deletes (delete_object/_character/_animation) need a confirm=false-then-true
two-step; tileset/tile deletes take just the id.
create_map_object auto-deletes after 8h — download immediately.
General
get_balance is the live source of truth for budget (not the REST balance command).
Backblaze rotation/B2 URLs are permanent and public — safe to reference while iterating.
Always check status with get_* before downloading; jobs are auto-cleaned after completion (404).
