A brand-new manim-render tool cap. The agent writes a Manim scene; the cap renders it into precise mathematical animation — real LaTeX equations that move, a secant sweeping to its tangent. The 3Blue1Brown look, from one prompt. The precise STEM tier of doc-to-animation, where HTML can't go.
The agent writes a Manim Scene; manim-render executes it. The equation is LaTeX (f'(x)=\lim_{h\to0}\frac{f(x+h)-f(x)}{h}), the tangent line genuinely sweeps to its limit, and f'(x)=2x resolves at the end — all exact, because it's computed, not generated.
class Derivative(Scene): def construct(self): ax = Axes(x_range=[-3,3], y_range=[-.5,6]) curve = ax.plot(lambda x: x**2) # f(x)=x² defn = MathTex(r"f'(x)=\lim_{h\to 0}\frac{f(x+h)-f(x)}{h}") h = ValueTracker(1.8) # secant gap → 0 self.play(h.animate.set_value(0.05)) # secant → tangent self.play(Write(MathTex(r"f'(x)=2x"))) # the result
manim-render (kind:"tool") is a Manim Community container on the Livepeer tool-host — sibling to hyperframes-render. It takes an LLM-authored Scene (Python), renders with manim render (minimal LaTeX baked in for MathTex), and returns an MP4. A static safety scan rejects any scene that reaches for the filesystem, network, or process control — scenes are pure animation. Where hyperframes-render owns UI/text/charts, manim-render owns math that has to be exactly right and move.
f'(x)=2x are typeset by LaTeX, not approximated by an image model.ValueTracker; that's what makes the idea click.f(x) or a range and re-render; it's code. Cheap (~$0.60).