How
Code generation only: an agent writes TypeScript scenes and can edit the meta JSON of time events. No MCP server or control API found. For Manim, community MCP servers (e.g. paulnegz/manim-mcp) wrap LLM text-to-scene generation and rendering.
Motion Canvas
TypeScript generator animations with a timeline editor for voice-over sync.
motioncanvas.io ↗repo ↗docs ↗Time events docs ↗Revideo (MIT fork for programmatic rendering) ↗Manim Community ↗manim-mcp (community text-to-animation MCP) ↗
- license
- MIT. Manim (3b1b and Community editions) is also MIT.
- language
- TypeScript
- backing
- Individual: Jacob (aarthificial, YouTube creator) wrote ~645 of the commits; community contributors; no company or foundation.
- stars
- 19,263 (as of 2026-10-09)
- downloads
- 5,566/week npm @motion-canvas/core (api.npmjs.org, 2026-10-01..07). For comparison: manim 45,052/week PyPI (pypistats).
- latest release
- v3.17.2 (stable); v3.18.0-alpha.0, 2024-12-14
- first release
- 2023-02
borrow
Steal the time-event model for tour narration; the tool itself targets video, not live terminal UIs.
Borrow from it: take a specific idea, API or format.
the steelman: the best honest case for it
Motion Canvas solved the hardest practical problem in programmatic explainers: keeping animation in sync with narration without hard-coded timestamps. Scenes stay honest code (typed, versioned, reviewable), while timing is a separate, draggable JSON layer; re-record the voice-over and you nudge one pill and everything after it follows. Because scenes are deterministic generators, every frame is reproducible and the editor can seek anywhere instantly. For an author who thinks in code but teaches with a voice, it is the most elegant split of 'what' from 'when' in any tool, and its vector rendering looks like hand-made motion design. Manim, its larger cousin, adds a decade of mathematical explainers and an active community.
scores
| UI as data | 1 scene 5 | Scenes are code; only time-event timings live in JSON. |
|---|---|---|
| Agent can drive it | 1 scene 5 | |
| Agent can see it | 2 scene 5 | Rendered frames are images an agent could inspect. |
| Small, targeted edits | 1 scene 5 | |
| Time travel | 3 scene 0 | Scrubbable deterministic timeline in the editor; not an agent API. |
| Live data rate | 0 scene 5 | |
| Terminal native | 0 scene 5 | |
| Pixel graphics | 5 scene 4 | |
| Teaching | 4 scene 3 | Built for narrated explainer videos with voice-over sync, but output is passive video. |
| Maturity | 3 scene 1 | Used in production by explainer creators; development stalled. |
| Openness | 4 scene 4 | MIT, single-maintainer control. |
Motion Canvas scene now scene planned
how agents use it
Wiring it in
Generate scene files, run the Vite editor or a render, inspect output frames.
Seeing the result
Rendered frames/video only; no structured scene-graph readback API for an external agent.
Small edits
Edit source or the time-events JSON; hot reload picks it up. No id-addressed live patching.
History
Deterministic timeline means any frame can be re-derived; scrubbing is human-facing in the editor.
In short
Great for agents that write code and render; nothing for driving a live instance.
architecture
Scenes are TypeScript generator functions: each `yield*` advances time, so an animation is a deterministic program over a frame clock. A Vite-based editor previews in real time and lets you scrub a timeline. Time events (`waitUntil('event')`, `useDuration('event')`) are named pauses whose timing is set by dragging pills in the editor and stored in a JSON meta file beside the scene, so audio sync is data, not code. Rendering is 2D canvas (vector shapes, text, LaTeX, code blocks with diff animations); export is image sequences or video via plugins.
- Determinism: frame N is a pure function of the scene program plus meta, so seeking anywhere re-runs the generator to that frame.
- Time events separate 'what happens' (code) from 'when' (editor-adjusted JSON), and shifting one event ripples later events.
- Signals (reactive values) bind properties to each other.
- Manim (Python) is the larger peer: renders to video offline, Community edition actively released; no live editor timeline of the same kind.
performance
No published numbers. Rendering is offline-to-video; live preview speed is not benchmarked publicly.
adoption and upkeep
Adoption
Well-known among explainer-video creators (19k stars) but small install base; Manim (94.8k stars 3b1b, 41.4k Community) dominates the code-animation niche.
Used by: aarthificial's YouTube explainers, Revideo (fork, 4.1k stars)
Maintainability
Effectively dormant for ~20 months; usable but not moving. Manim Community, by contrast, released v0.22.0 on 2026-10-08.
Releases: Stalled: v3.17.0 2024-08-13, v3.17.1 2024-08-17, v3.17.2 2024-12-14, v3.18.0-alpha.0 2025-02-16; none since. · Contributors: ~98 (GitHub contributors incl. anonymous, via Link header) · Recent: 0 commits since 2026-07-11; the only 2025-2026 commit after Feb 2025 is a docs-domain update (2026-07-02). · Bus factor: low: one author wrote the large majority of commits and activity stopped in Feb 2025.
weaknesses
- Dormant: no release since 2025-02-16 alpha; no commits since 2026-07-11 (GitHub API).
- Bus factor of one: ~645 of the author's commits vs a handful from others.
- Output is passive video: a viewer cannot take over or interact.
- No agent control surface or structured readback; Revideo forked it partly to add programmatic rendering.
- Small install base (~5.6k npm downloads/week).
and thc-scene
Overlap
Narrated walkthroughs driven by a timeline; deterministic replay of a pure program over time.
What scene would be reinventing
Syncing narration to visual steps; Motion Canvas's time events already model this cleanly (named waits, adjustable durations, ripple editing).
The gap it leaves
A live, interactive, agent-driven UI the learner can take over in a terminal; Motion Canvas produces video and has no runtime control.
What to borrow
- Named time events stored as data beside the tour: `waitUntil(name)` with timings editable without touching the script.
- Ripple semantics: moving one event shifts all later events unless held.
- Deterministic seek by re-running from the start (or a checkpoint) to frame N.
sources
- Motion Canvas docs: Introduction motioncanvas.io
- Motion Canvas docs: Time Events motioncanvas.io
- GitHub API: motion-canvas/motion-canvas api.github.com
- Meta file timeEvents example (Ahmad Rosid tutorial) ahmadrosid.com
- HN: Revideo fork discussion news.ycombinator.com
- npm downloads @motion-canvas/core api.npmjs.org
- pypistats manim pypistats.org
- paulnegz/manim-mcp github.com