Manim

Manim

Generate a short mathematical / diagrammatic animation video for embedding in an HTML page. This is a thin orchestration layer: Manim is the externally provisioned runtime — it is never vendored or reimplemented here. The driver script composes Manim's programmatic render path; uv provisions the Python stack; Cairo + Pango come from the system.

Respond in the user's language.

Prerequisites

Preflight

Verify the system libraries before authoring (mirrors a which-style gate). The first manim checkhealth run also provisions the Manim stack into uv's cache:

which pkg-config && pkg-config --exists cairo pango && echo "cairo/pango OK"
uvx --from 'manim==0.20.*' manim checkhealth

manim checkhealth verifies Cairo/Pango and warns if LaTeX is absent — that warning is expected and fine; LaTeX is opt-in (see below). If pkg-config is missing, the first render's wheel build for pycairo will fail — install it now.

Workflow

Phase 1: Author the Scene

Translate the user's request into a Scene subclass in a .py file. Keep it short (a few seconds — these are loopable web clips) and silent (Manim output has no audio track, which is what muted-autoplay wants).

from manim import *

class MyScene(Scene):
    def construct(self):
        axes = Axes(x_range=[-3, 3], y_range=[-2, 2])
        graph = axes.plot(lambda x: x**2 / 2, color=BLUE)
        self.play(Create(axes), run_time=1)
        self.play(Create(graph), run_time=1.5)
        self.wait(0.5)

Authoring guidance:

• One Scene subclass per request is simplest; the driver auto-selects the sole one (pass --scene NAME when a file defines several).
• No-LaTeX text path: use Text(...) / MarkupText(...) (Pango-rendered) for labels. Reach for MathTex(...) / Tex(...) only for real equations — that path invokes LaTeX and requires the opt-in toolchain.
• Common building blocks: Axes/NumberPlane + .plot() for graphs, geometric mobjects (Circle, Square, Polygon, Arrow), ParametricFunction for curves, ThreeDScene + ThreeDAxes for 3D. Animate with self.play(Create(...), Transform(a, b), FadeIn(...), ...) and pace with run_time / self.wait().

Write the scene to a working file (e.g. under the scratchpad), not into the repo.

Phase 2: Render (headless)

Run the driver. It provisions Manim on first call, renders with the default software Cairo renderer (no display needed), moves the file to a known path, and prints the absolute path + embed snippet.

uv run ${CLAUDE_PLUGIN_ROOT}/skills/manim/scripts/render.py <scene.py> [flags]

Examples:

# Default muted-loop mp4
uv run ${CLAUDE_PLUGIN_ROOT}/skills/manim/scripts/render.py /tmp/scene.py

# Transparent overlay (auto-webm)
uv run ${CLAUDE_PLUGIN_ROOT}/skills/manim/scripts/render.py /tmp/scene.py -t

# Equation scene (LaTeX preflight) at high quality
uv run ${CLAUDE_PLUGIN_ROOT}/skills/manim/scripts/render.py /tmp/scene.py --latex -q high

Phase 3: Return

Report two things to the user:

1. The absolute path to the rendered file (the MANIM_OUTPUT_PATH= line).
2. The embed snippet, ready to paste:
   • mp4 / webm → <video autoplay muted loop playsinline src="..."></video>
   • gif → <img src="..." loading="lazy">

The src is the local absolute path (previewable via file://); tell the user to swap it for their served URL when hosting. Do not build any further HTML-page integration — emitting the snippet is the whole output contract.

Notes

• Format choice: mp4 is the safe default (universal browser support, small). Use webm only for transparency. Use gif for very short inline loops where a <video> element is overkill — but gifs are larger and lower quality.
• First render is slow — uv resolves and caches the Manim stack (numpy, scipy, pycairo, manimpango, av, ...). Subsequent renders are fast.
• LaTeX stays off by default. Equations are the only trigger; everything else (text via Pango, shapes, graphs, 3D) renders without it.