agent-sprite-forge: one approved still, then a measured pipeline to a game-ready sprite set
Agent Skill for generating 2D sprite sheets and map, transparent PNG frames, and animated GIFs from prompts.
At a glance
- What is it?
- agent-sprite-forge is a Python agent skill for Claude Code, Codex and Grok that turns one approved character still into keyed, looped, retimed actions exported for Godot, Aseprite or a web game. The pipeline is checked by scripts, but it leans on external image and video routes you must already have.
- Who is it for?
- Use it if you already run Claude Code, Codex or Grok and want one character turned into several engine-ready actions with measured loops and timings. Skip it if you need a standalone sprite editor or have no image and video route to point it at.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- 0x0funky/agent-sprite-forge5,754 ★ · +1,576
| Repository | 2026-09-23 | 2026-11-15 |
|---|---|---|
| 0x0funky/agent-sprite-forge | 4.2k | 5.8k |
1,576 Stars in the last 53 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Everything downstream hangs on one approved master still
The workflow starts with a single image. master_still.py writes the prompt, generates takes, fixes a near-miss by edit and approves one still, which is recorded in master.json. Only after that does sprite_set.py make one image-to-video clip per action from the approved still.
The consequence is that the still is the single point of failure for the whole set. Six actions in the project's own example all inherit the same character design, so a flaw you approve at step one is repeated in every clip, and the checks that compare each take against the master cannot catch a mistake the master itself contains. The example in the project shows take 1 of 3 being approved, which suggests you should expect to spend a few generations here before the rest of the pipeline is worth running.
The pipeline also fixes one body scale per character from the master, so every action of a set is sized consistently rather than per clip.
Installing for Claude Code leaves the Python dependencies to you
The three hosts install differently, and the differences matter. For Claude Code the commands are:
claude plugin marketplace add 0x0funky/agent-sprite-forge --sparse .claude-plugin skills
claude plugin install agent-sprite-forge@agent-sprite-forge
python -m pip install "numpy>=1.26" "Pillow>=10.1" "scipy>=1.11"The --sparse flag checks out only the plugin manifest and the skills, not the showcase media. The plugin install does not pull the Python packages, so the last line is a separate step, and skipping it means the deterministic tools have nothing to import. The same three pins appear in requirements.txt: numpy>=1.26, Pillow>=10.1 and scipy>=1.11.
For Codex you clone the repository, run python -m pip install -r requirements.txt, then python tools/install_skills.py --apply --host codex. Grok uses the same script with --host grok, which writes to ~/.grok/skills. Start a new agent session afterwards or the skills will not load.
The installer backs up what it overwrites and can report drift
tools/install_skills.py is more careful than a plain copy. With --apply it backs up an installed copy and writes a sha256 manifest. Later, python tools/install_skills.py --check --host codex reports drift between what is installed and what the repository holds.
Other targets exist: --host agents writes ~/.agents/skills, and --dest accepts any skills folder. That covers hosts that are not named in the project, but the project only names Claude Code, Codex and Grok as supported agents, so a custom folder is your own responsibility to wire up.
The practical result is that updating is a re-run of the same command, and the manifest gives you a way to notice a hand-edited skill. What the manifest does not do is merge your edits. If you changed an installed skill, --apply replaces it, with a backup as your only safety net.
route_media.py picks a route by what you are signed into
Every image and every clip goes through one command, route_media.py. It tries your API key first if you configured one, covering OpenAI, Google Gemini, xAI, BytePlus and fal.ai. If there is no key it uses your signed-in Codex or Grok CLI. Code art is the last resort.
In the project's example the master still came from the local Codex CLI through image_gen, and the clips came from the Grok image-to-video local CLI. So the headline result depends on at least one generation route being available on your machine, and the quality of the sprites is bounded by whichever route wins the fallback order.
The cost of this design is that the same prompt can take a different route on a different machine. If you configure no key and are not signed in to either CLI, you land on code art, which is not the image-to-video path shown in the example frames. Decide the route deliberately before comparing output with the published examples.
Twelve gates decide whether a take is kept or retaken
sprite_set.py does not accept a clip because it exists. It gates every take on area, feet, identity, zoom, turning around, edges, background, extra objects, motion, end pose, colour bleed and timing. A failing take is retaken with fix clauses added to the prompt.
The gate thresholds were recalibrated on 42 labelled live takes. In the project's table, good takes wrongly rejected went from 18 of 24 to 0 of 24. That is a measurement on one labelled set of takes, so it tells you the old gate was too strict on that set, not that the new gate is right on your character.
The consequence for you is time and generation cost. Every rejected take is another image-to-video call through your route, and a character that repeatedly fails one gate, such as turning around, will keep retaking. The pipeline reduces manual review, but it does not promise a set on the first pass.
The before and after numbers come from a single live run
The 0.4 comparison table is specific. For the pixel fox at 48x64 with 8 frames, colours per frame fall from 594-709 to 16-19, alpha levels from 199 to 2, and mean lone pixels per frame from 732 to 75. The soft-matte keyer takes purple fringe from 0.704 to 0, leaked key pixels over 145 frames from 7,780 to 0, and keying time from 271 s to 44 s. The attack one-shot drops from 4.1 s to 0.7 s, and colour lock cuts hue flips per frame pair from 352 to 170.
Treat these as one run's results. The frames in the project come from the 2026-10-06 live run, and the project points to docs/validation-2026-10-06.md as the record, weak points included. A fox sprite and a Grok clip do not guarantee the same gains for a different art style or a different route.
Colour lock halves hue flips, not eliminates them, so some shimmer on fast limbs can remain.
HD is the default and the pixel finish is opt-in
The finish stage defaults to HD. Crisp pixel finish happens only on request, and it adds a crisp downscale, a spaced palette of at most 32 colours, a selective outline and binary alpha.
If you want pixel art and forget to ask, you get HD frames with soft alpha, which will not match a pixel-art game's look and will need a second pass. The reverse also holds: the palette cap of 32 colours means a pixel finish cannot preserve a richly shaded character.
One-shot actions need extra care because image-to-video plays a quick attack as slow motion. retime --auto-oneshot cuts the frozen holds and pins the hit, and the example attack lands at 292 ms. Loops such as run are picked by measurement, and the scene example measures run speed from the contact foot so feet do not slide. If your character has no clear contact foot, expect that measurement to give you less.
Video needs ffmpeg 5.1+ and the engine exports are all-or-nothing per format
Video work needs ffmpeg 5.1 or newer on PATH. Without it the clip stages cannot run, and an older ffmpeg is outside the stated requirement.
The export list is concrete. animation.json 3.0 comes with a PNG atlas, plus VP9-alpha WebM and packed-alpha MP4 that are decoded and verified. Godot gets SpriteFrames and Sprite3D, Aseprite gets JSON, and maps go to Tiled, Godot and LDtk. The project also notes Windows hardening for Codex sandbox run folders, locked-file cleanup and cp1252 and cp950 consoles.
The release history is short: v0.4.0, titled The Big Upgrade, was tagged on 2026-10-06, the same day as the last push. That makes this a young pipeline whose repository is current, but with only one tagged release listed. License is MIT and the repository is not archived.
Editorial conclusion
Use it if you already run Claude Code, Codex or Grok and want one character turned into several engine-ready actions with measured loops and timings. Skip it if you need a standalone sprite editor or have no image and video route to point it at. Before relying on it, read docs/validation-2026-10-06.md for the weak points, and check that ffmpeg 5.1+ is on PATH.
Frequently asked questions
What do sprites do in agent-sprite-forge?
Here a sprite set is the per-action frames of one character, keyed with transparency, looped or retimed, and packaged as animation.json, Godot SpriteFrames, Sprite3D or Aseprite JSON.
What is the best program for creating sprite art?
agent-sprite-forge does not rank programs. It is an agent skill that generates one master still, then clips per action, and exports Aseprite JSON among other formats.
What does sprite mean in gaming?
In this project a sprite is a character animation frame or action exported as transparent PNG frames, GIFs or engine assets for Godot, Aseprite or a web game.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/0x0funky-agent-sprite-forge)Star history chart for your README
[](https://hysenlabs.com/projects/0x0funky-agent-sprite-forge)Was this useful?