Make your own pet¶

It looks like you want a pet that isn't a paperclip. Rude, but fair. Here's how.

Everything you need to make a pet is in this repository, and none of it is secret. There are two routes:

Route A: remix Clippy Pet¶

The normalized per-state frames live in source/frames/<state>/NN.png (MIT-licensed; the unofficial notice covers the likeness). Recolour, re-shade, swap the eyes, add a hat. Then reassemble:

Keep every frame 192 × 208 with a transparent background and the character's baseline where it is.
Lay frames left-to-right into rows in this order: idle, running-right, running-left, waving, jumping, failed, waiting, running, review, then the two look rows (see the contract).
Export a lossless RGBA WebP at 1536 × 2288.
Update pet.json: new id (lowercase, hyphens), displayName, description.
Run the validator; if you changed frame counts, adjust expected_used in scripts/validate.py to match your rows.
Drop the two files in ~/.codex/pets/<your-id>/, reload, admire.

Assemble an atlas with Pillow (Python)

from pathlib import Path
from PIL import Image

ROWS = ["idle", "running-right", "running-left", "waving", "jumping",
        "failed", "waiting", "running", "review", "look-a", "look-b"]
W, H, COLS = 192, 208, 8
atlas = Image.new("RGBA", (W * COLS, H * len(ROWS)), (0, 0, 0, 0))
for r, state in enumerate(ROWS):
    for c, frame in enumerate(sorted(Path(f"frames/{state}").glob("*.png"))[:COLS]):
        atlas.alpha_composite(Image.open(frame).convert("RGBA"), (c * W, r * H))
atlas.save("spritesheet.webp", lossless=True, quality=100, method=6)

Route B: hatch something new¶

Start from the contract and the design notes. The advice that mattered most for Clippy Pet, in order:

Design the silhouette at 64 px first. If it doesn't read in flat black at that size, no amount of shading will save it.
Put the expression in the eyes. They're the only feature guaranteed to survive downscaling.
Author loops, not clips. Last frame flows into first; the host chooses timing.
Mirror poses, not pixels for left/right so the light doesn't flip.
Do the look directions last, and test them blind. Shuffle the sixteen frames, hide the labels, and ask someone which way each looks. Ours got 13/16 clean on the first honest pass and we published the three warnings.
Ship the receipts. A contact sheet and a validator make your pet reviewable, and reviewable pets get shared.

OpenAI documents the pets feature and its file format on the ChatGPT help site; if their contract changes, that page wins over anything here.

Reuse the packaging, too¶

Because Clippy Pet's installers are payload-agnostic, you can fork the repo, replace pet.json/spritesheet.webp, change the name in VERSION, pet.json, packaging/linux/nfpm.yaml, and packaging/bin/clippy-pet (PET_ID), and inherit the one-line installer, the .deb/.rpm/.apk/Arch builds, the macOS .dmg, checksums, signatures, and this documentation site. The packaging runbook is the map.

Show it off¶

Made something? Post it in Show and tell with a GIF. If it's a Clippy Pet variant you'd like merged, read Submit a variant.

Submit a variant