AI video generators and AI-written HTML animations can look impressive, but they are hard for a person to adjust piece by piece. Fixing one word or slowing down one element can mean generating the whole thing again. Connecting Claude Code to Adobe After Effects changes that. Through the Model Context Protocol (MCP), an open standard that lets AI models connect to outside tools and data, Claude can create real text layers, shape layers and keyframes inside After Effects. What comes out is a project file that a person can open, adjust and pass along, not a flattened video. This article covers the setup and two tests: rebuilding an existing motion graphic from a 6.7-second clip, and building a 20-second product launch video with no reference at all.
Why an editable project matters
After Effects is widely used industry software for motion graphics, UI mockups, animated titles, product explainers and diagrams. A motion graphic in After Effects is assembled from separate text, shape and background layers, and keyframes on a timeline decide when and how each layer moves. A keyframe records the value of a property, such as position, at a specific point in time.
Because Claude works inside that structure, everything it produces stays editable. You can ask in plain language for a new color, position or line of text without rebuilding anything, or make the change yourself: wording sits in a text layer, and speed comes from the spacing between keyframes. For a one-person operation, that allows a practical split of labor. AI produces a working first draft, and a professional editor receives the native .aep project file for the final polish.
What you need
The core requirements are short: a Claude subscription and a copy of Adobe After Effects installed on your computer. No third-party video platform has to sit in the middle.
In this workflow, Claude Code runs inside the Claude desktop app with the Claude Opus 5.5 model. It also works from the terminal, which may be slightly faster, but the desktop chat interface appears easier to follow. Because After Effects runs on your own machine, set the Claude Code session to Local rather than Remote.
The bridge between the two programs is an openly available After Effects MCP server. It hands Claude a set of functions to call, such as get_project_summary for reading the project state, create_shape_layer and create_text_layer for adding layers, and add_keyframe and set_temporal_ease for animating them. Claude Code runs those tool calls, along with any file operations, directly on your computer.
Three steps before the first animation
- Check the connection. Ask Claude to run
check_setupandget_project_summary. Together they confirm that Claude is attached to the open project and that the project starts out empty. - Set up folders. Create one working folder with separate subfolders for projects, references and renders.
- Save versions. After every successful generation step, save the project under a new name, such as v1, then v2.
The third step matters most. A request to tweak one small element can lead the model to rewrite the entire composition and break a layout that was already working. With earlier versions on disk, you can roll back without losing work.
▲ MCP connection and saved project versions
Test one: rebuilding a graphic from a 6.7-second clip
The first test recreates an existing motion graphic from a tech education company’s video. Its message is simple: infrastructure has to be in place before an AI layer goes on top. A dark base block labeled “01 Infrastructure” sits at the bottom, an outlined, blueprint-style block labeled “02 AI” floats above it, and an accent tag and a soft floor shadow complete the design.
The 6.7-second MP4 of the original goes straight into Claude Code, together with a prompt asking for separate text and shape layers and for keyframes that follow the motion in the clip. Claude then samples the footage at several moments, such as 0.6, 2.0 and 4.5 seconds, to work out which layer enters first and how fast each one travels. From that analysis it builds the composition, the After Effects container that holds a scene’s layers.
This is why a clip makes a better reference than a screenshot. A single still carries no timing information, so it cannot reveal entrance order or speed. Several sampled moments give the model enough sense of time to match both timing and easing, the way a movement speeds up and slows down.
Placed side by side with the original, the rebuilt graphic closely follows the layer placement, type size, block proportions and easing. The difference is that every element of Claude’s version remains a layer you can change.
Revising without breaking what works
Revisions go best as a short, numbered list. A typical round looks like this:
- Change the peach panel to blue.
- Replace the “first” label with “Foundation.”
- Slow the lower block’s movement and bring the AI block in right after it.
- Keep every other element exactly where it is.
Locking down everything you want to keep makes it easy to confirm what changed. Packing dozens of edits into one prompt tends to confuse the model and hides what actually moved, which can trap you in an endless cycle of reviewing and re-requesting. Order matters as well. Settle the overall layout and the hierarchy of the blocks first, and leave fine visual polish for later rounds.
Inside the project file
Opened in After Effects 2026, the revised v2 project shows separate, named layers such as CARD, CAM, TOWER, TITLE and BASE_Plate. Next to the shape and text layers sit null rigs, invisible control layers that move groups of layers together, and track mattes, which use one layer to define the visible area of another.
Small wording changes do not need another prompt. Select the text layer and type over it, for example turning “Engineering” into “Design,” and the update is done without asking the AI to rebuild anything.
Timing is just as easy to reach. Each keyframe stores a time and a value, so the distance and the time between two keyframes set a layer’s speed and path. In one example, a position value travels from 780 to 420 between 1.0 and 2.5 seconds. Dragging the end keyframe farther along the timeline stretches that move, and the entrance becomes slower and more gradual.
Test two: a 20-second launch video from scratch
The second test is harder: create a new piece with nothing to copy. The subject is a launch video for a SaaS product, and it makes a useful benchmark because a professional motion graphics designer had already produced one for the same product.
The prompt includes the product specifications, brand guidelines and visual reference documents, along with the order of the scenes. The story moves through four stages, Idea, Angle, Script and Package, where the package covers a title, a thumbnail direction and a description. A single running example, making a video about Claude and After Effects, carries through every scene so the context never resets. The target length is about 20 seconds, which gives each scene room to breathe without cramming in extra feature demos.
| Setting | Value |
|---|---|
| Runtime | About 20 seconds |
| Resolution | 1920x1080 |
| Frame rate | 30 fps |
| Scenes | Idea, Angle and Script, Package, Wordmark |
| Inputs | Product specs, brand guidelines, visual references, scene order |
Before playing the full render, check a still frame from each scene for legible type and a consistent brand look.
After the first draft, the revisions focus on flow. In the package scene, the script card stays in view while the title, thumbnail direction and description panels animate in around it, which holds the scene together. The ending gets a closing call to action next to the product’s logo and web address, with enough hold time at the end to read them.
The resulting project keeps each scene’s composition in its own subfolder, complete with rigs, slider controls and fade expressions, the short scripts that calculate property values in After Effects. Because the scenes are modular, an editor can adjust a single camera move, sketch line or background card without regenerating the whole sequence.
▲ Human designer craft versus an AI first draft
How it compares with a human designer
Set beside the professional designer’s version, with layout, typography and motion feel compared directly, the gap is easy to see. The human-made piece carries a distinct artistic nuance, and Claude does not yet appear to fully match the subtle craft of a skilled motion designer.
Still, the AI version is a viable, fully editable working project, and a more competent one than many people might expect. Its realistic role may be to deliver a fast first draft that a person then refines, which can speed up production considerably.
Getting better results
Left to make every visual decision on its own, the model turns motion design into a lottery. Clear creative direction, a structured scene outline and brand assets lead to markedly better results that are also easier to repeat. The cleanest output tends to come from recreating concrete material such as UI mockups, real screenshots and screen recordings, while inventing abstract artwork from scratch appears to be a weaker area.
A practical way to start:
- Get a Claude subscription and After Effects, open a Claude Code session in Local mode, and connect the After Effects MCP server.
- Run the connection checks, then set up separate folders for projects, references and renders.
- When recreating something, give Claude a short clip instead of a still, and spell out the scene order and total length.
- Ask for small changes you can verify, and save a new version after each successful step.
- Make simple text and timing tweaks yourself in After Effects.
- Keep the final render (.mov) together with the editable project (.aep) so you or an editor can change wording, easing or timing later.
The real value of this setup appears to be that AI output stops being a one-shot render and becomes a working file that people can keep improving.