Usage in Chat

Ask for something in a session (for example "draw a unit circle with Manim"); the AI submits a piece of Manim Python through the manim_card tool, and a card appears in the chat stream. Before the tool returns, the code is prechecked: syntax errors, undefined names and unsupported constructs come back to the model, which corrects them within the same turn.

Playback rules

Several cards may exist in the chat stream at once, and playing all of them in full is expensive, so full playback happens only in these cases:

Case	Behaviour
A newly generated card first enters the viewport	Plays through once
Any later remount (switching back, refreshing, scrolling away and back)	Lands directly on the final frame, no replay
Cards in historical sessions	Do not play; show the final frame
Several cards produced within one step	Only the last one plays in full
Scrolled out of the viewport	The canvas is kept; when more than 8 canvases are active they are reclaimed from farthest to nearest relative to the reading position

The playback chance is consumed exactly once

Scrolling out of the viewport after entering the playback queue, or skipping midway, spends the chance — viewing the card again shows the final frame, not a replay from the beginning.

Manual playback

The button in the card's top bar has two states:

• ▶: play in full. Takes precedence over all the automatic rules above.
• ⏩: jump straight to the final state.

Interrupted cards (the turn was stopped) do not render by default; press ▶ to start. Every canvas on a page shares one execution queue, so two canvases never play at the same time.

Viewing the code

The Python source can be read on the card and copied out to the AI for further adjustment.

When nothing can be drawn

On a render failure the card shows manim's original error message at the bottom. If a hard environment condition really is unmet, the card adds one actionable line — the same verdict the "Browser runtime" card in the Settings panel reads. Per-item states are in the Settings Panel.