Can Google Antigravity Create and Edit Videos?

Yes. Google Antigravity can create and edit videos by working with a video tool such as CoAnimator. Antigravity handles the instructions and project edits; CoAnimator provides the stage, timeline, and local video export. Together, they let you describe a product demo, review the scenes, and revise the result in an editable project.

That is useful when a video needs more than an attractive opening shot. A software demo has to show the correct interface, explain a real action, and remain fixable when the product changes. You need control over the second version as much as the first.

Gemini 4 Argon adds a timely reason to explore this workflow. Google began rolling it out in Antigravity on October 13, 2026, and creators are already sharing animation experiments. This guide covers what those capabilities mean in practice, how to connect Antigravity to CoAnimator, and how to take a video from brief to reviewed MP4. The setup follows the current documentation. The saved-view demo below is an illustrative production brief that you can adapt to your own assets.

What Antigravity does in a video workflow

Antigravity is Google’s agent development environment. Its agents can work with files and tools to carry out a task. In this workflow, the task is a video project: reading the brief, writing scene instructions, adjusting the sequence, and invoking the renderer.

CoAnimator makes that project visible. You can inspect the scene on the stage, see how it fits the timeline, and request a change while keeping the underlying source. The CoAnimator creation guide describes this progression from product research and script to individual scenes.

The responsibilities are easier to understand when separated:

This supports several kinds of work: an animated feature explanation, a launch video using product screenshots, or a tutorial built around recorded footage. The starting material determines what the agent edits. An animated title remains an editable element; a title already baked into an imported recording is part of that recording’s pixels.

The model influences how well the agent interprets your brief, but the source material determines what it can change. That is the useful context for the latest Argon release.

Where Gemini 4 Argon fits, with Rege’s animation example

Google’s Gemini 4 Argon announcement describes its rollout to the Antigravity desktop app and CLI for Google AI Ultra subscribers and eligible Gemini Enterprise subscriptions. Check your account’s available models rather than assuming every Antigravity installation includes Argon.

Before installing anything, check which Antigravity product you are using: the Antigravity app and the standalone Antigravity IDE are different surfaces. Google’s getting-started documentation says Argon is not available in the standalone IDE and directs users to the app, CLI, or IDE extensions. The CoAnimator workflow below uses the CLI.

On October 14, Rege shared an animation experiment with Argon, writing that he had been experimenting with the model and exploring animation ideas. Watch the post for a concrete example of the creative work being shared around the release:

The post shows an experiment; it does not document the full toolchain, number of revisions, or production time. It is a useful creative reference, rather than a reproducible benchmark for this tutorial.

When watching a reference like this, identify something you can direct: how a shape enters, where the eye goes, or how long a composition holds. “Use a restrained reveal, then pause on the product result” is actionable. “Make it look like that” leaves the agent to guess which qualities you liked.

You can bring those observations into your own brief. First, though, make sure Antigravity can reach the project and the renderer; a model selection alone does not establish that connection.

Set up Antigravity with CoAnimator

The setup has three separate pieces: the agent, the video application, and the instructions that teach the agent how to work with it. Prepare all three before asking for a finished video.

Install the studio and rendering tools

Download CoAnimator, open it, and create a blank project or choose a suitable template. Use its media-tool setup to prepare FFmpeg and FFprobe. Then open Settings → General → Command line & agent skill and install the coa command.

The CoAnimator CLI reference documents these dependencies. A useful first check is to run the following in a fresh terminal on macOS or Linux:

coa --help
ffmpeg -version
ffprobe -version

In Windows PowerShell, use coa.cmd --help for the first command. Opening a new terminal matters because an already-running session may not know the newly installed command location.

Install Antigravity CLI and sign in

Follow Google’s installation instructions. The command you will use is agy. Run it once to complete authentication, and check the models available to your account.

In CoAnimator, select Antigravity from the agent picker. The documented integration opens a terminal tab in the project directory, giving you Antigravity’s own interface and history. Approvals happen there, rather than through CoAnimator’s chat-agent approval controls. See the Antigravity CLI connection guide.

Use the same local machine for this walkthrough. If you later move execution to another computer, that environment also needs the project media and rendering tools.

Add the CoAnimator skill

Download the SKILL.md linked from CoAnimator’s skills page and put it inside this project-relative folder:

.agents/skills/coa-cli/SKILL.md

Google documents .agents/skills/ as a supported workspace skill location. Its skills reference also lists global locations, which differ between the app and CLI. A project-local copy keeps this setup easier to inspect.

The skill supplies guidance, not software. It does not install coa, sign you into Google, or configure media providers. Start a new agent session and ask it to read the skill and relevant bundled documentation before editing.

Finish setup with a small request: “Inspect this CoAnimator project, confirm which tools are available, and explain how you would render it. Do not change the scenes yet.” Once the agent identifies the project correctly, you can move from installation to creative decisions.

Plan a product video around one visible outcome

For a first project, choose a feature whose value can be shown on screen. “Our platform improves productivity” is too broad to direct. “Save a filtered view so the team can return to the same tasks” gives the video a starting state, an action, and a result.

The example here is a hypothetical 35-second demo for a project-management product. It teaches existing users to save a view. The timings and filenames are illustrative, so replace them with your own material.

Before prompting, gather a current recording of the action, a clean capture of the result, the logo, and approved typography and colors. Use a test workspace with realistic sample data. A screen filled with “test123” distracts from the feature; real customer details introduce an unnecessary problem. The recording should show what the shipped product actually does.

Keep those files together and name their purpose clearly. A folder containing save-view-recording.mp4, saved-view-result.png, and brand-notes.md gives the agent better context than several unexplained screenshots. If one image is only a style reference, say so; otherwise it may be treated as a screen to reproduce.

Write a brief that produces a reviewable plan

Ask for the plan before the animation. It is easier to correct the emphasis or remove an unsupported claim in a scene plan than in a finished animation.

Plan a 35-second CoAnimator video for existing users of our project-management app. Teach one task: save a filtered view and reopen it later.

Use save-view-recording.mp4 for the real action and saved-view-result.png for the final hold. Read brand-notes.md for fonts and colors. Use a 16:9 canvas and make the story understandable without sound.

Show the starting list, applying the filter, saving the view, and returning to it. Finish with “Try a saved view.” Do not invent interface controls, customer results, or time-saving statistics.

Return a scene plan with timings, on-screen copy, asset choices, and any missing information. Wait for approval before building scenes.

The prompt establishes what the video must prove. It also makes uncertainty visible: if the recording does not show the saved view reopening, the agent should flag the gap instead of fabricating a successful result.

Our product-video briefing guide expands on audience and asset direction. Although its examples use Claude Code, the same editorial questions apply here: who is watching, what should they understand, and which screen demonstrates it?

Turn the plan into a sequence viewers can follow

For this example, a practical starting structure is:

1. 0–5 seconds: establish the task. Show the list and the need for a reusable view. Make clear which information the viewer wants to find.
2. 5–14 seconds: apply the filter. Let the viewer see the list change before introducing Save.
3. 14–23 seconds: save and name the view. Keep the chosen name visible so it is recognizable later.
4. 23–30 seconds: reopen the saved view. Show the result long enough to confirm the same filter is active.
5. 30–35 seconds: hold the result. Finish with a short next step that tells the viewer what to try.

These transitions carry meaning. The saved name links the creation step to the return visit; the held result confirms that the workflow succeeded. A decorative wipe cannot supply either connection.

If the action cannot fit comfortably, shorten the introduction before accelerating the important clicks. A tutorial should spend its time where someone might otherwise get lost. Once that sequence feels complete, visual design can make it easier to follow.

Build clear scenes before adding elaborate motion

Start with representative still frames: the opening, the main action, and the result. Ask Antigravity to establish composition and text hierarchy before animating every scene. This avoids polishing a sequence whose core screen is too small to read.

For the saved-view demo, the application should dominate the frame. Crop enough to make the filter and Save control legible, but retain context so viewers know where they are. Keep explanatory text outside the area where the cursor needs to move. Use the accent color to identify the current action rather than highlighting every panel.

A compact visual direction might be:

Use the approved dark palette and font. Keep the product screen large, with one short explanatory line per scene. Highlight only the control being used. Preserve enough of the sidebar to show location. Present three key frames before adding transitions.

Review those frames at the size the video will actually appear. A layout that looks excellent full-screen may become unreadable inside a help-center player. Test the longest label and the busiest screen now, when adjusting the composition is still straightforward.

Give motion a specific job

After approving the frames, ask for movement that directs attention. Let the screen settle before the pointer moves. Bring in the caption when its action becomes relevant. Hold after the result appears so viewers can compare the new state with what they saw earlier.

For example, a gentle crop toward the filter can remove distractions. A rapid zoom followed immediately by another zoom makes the same action harder to track. If the viewer must read a name, pause the movement while that name is on screen.

Ask Antigravity to build a rough cut with simple transitions first. Watch for continuity: does the cursor arrive from a believable location, does the selected view retain its name, and does the scene change preserve the viewer’s understanding of the task? Resolve those issues before adding background decoration or sound effects.

Fit narration and captions to the action

Write narration after the sequence is understandable visually. Describe what the viewer needs to notice rather than reading every label aloud. “Save this filter as a view” is more useful at the Save action than a broad sentence about team productivity.

If you use provider-generated voice or supporting media, check the available integration before making it essential to the brief. CoAnimator’s plugin documentation lists provider setup and notes that some actions need a supported agent environment. An enabled plugin does not establish that every terminal agent can operate every action.

You can also import approved audio. In either case, listen for clipped words, changes in volume, and narration that continues after the relevant screen disappears. Check captions against the final spoken wording. Avoid solving a long sentence by making the scene frantic; rewrite the sentence or give the action more time.

The same planning and review process also applies when you already have a recording. In that case, the next task is to select and refine the useful footage rather than construct every scene from scratch.

Edit an existing recording with a precise request

To edit footage, bring the recording into CoAnimator and tell Antigravity what should survive the cut. CoAnimator’s footage-editing guide covers media import and effects such as overlays and freeze-frames. Its current 0.7.2 documentation says desktop screen capture requires a separate recording tool.

Antigravity also has browser-recording capabilities, documented in Google’s feature overview. A browser capture can be useful source material, but it still needs selection and editing to become a clear product video. Watching every navigation step is rarely the best way to teach one feature.

Suppose you have a 90-second recording containing the saved-view workflow. The useful action starts at 00:18 and ends at 00:46. Instead of asking the agent to “make it punchier,” specify the boundaries and the information the edit must retain:

Create a short tutorial from save-view-recording.mp4. Use the action from source 00:18 to 00:46. Remove the opening navigation and pauses that do not explain the task.

Keep the filter choice, view name, Save click, and reopened result visible. Add a short title before the action and hold the result at the end. Preserve the original recording as a source asset.

Propose the cut points before applying them. Do not change playback speed during the filter selection or naming step. Do not add narration yet.

Distinguish source timestamps from output timestamps in later feedback. Once the first 18 seconds are removed, “fix the click at 00:22” can refer to two different moments. Include the scene name or a reference frame when there is any doubt.

A first cut may still need small adjustments. Watch the few seconds around each edit for a cursor jump, an abruptly disappearing menu, or a sentence cut mid-word. Then request a correction at that location: “In the naming step, hold the completed name for one second before the Save click; leave the other scenes unchanged.”

The same approach works when the product changes. Replace the outdated screen or action, update any related explanation, and compare the affected scene with the surrounding cut. Our guide to updating product videos covers that maintenance workflow in more detail.

Keep in mind that an imported MP4 does not reveal the editable layers of the project that produced it. You can trim, crop, or overlay the footage, but changing a baked-in label may require a replacement capture. Once the source is accurate and the edits feel natural, export the result for a final review outside the editor.

Export the video, then review the actual file

A preview helps you make decisions, but the exported file is what your audience will watch. In CoAnimator, select the intended timeline and export range before rendering. The export guide explains full-timeline, In/Out, and custom ranges, along with resolution and quality options. For a first delivery, start with the recommended Standard/Auto settings and change them only when the destination requires it.

For a repeatable terminal workflow, Antigravity can run the installed renderer. These examples run from the directory containing a project folder named saved-view-demo. Replace that folder with your actual project path. If your terminal is already inside the project, use . as the input instead:

# macOS or Linux
coa render ./saved-view-demo --out ./saved-view-demo.mp4

# Windows
coa.cmd render .\saved-view-demo --out .\saved-view-demo.mp4

The CoAnimator CLI reference documents supported inputs and render options. Have the agent report the output location and any errors. A command that was proposed in chat is not evidence that a render finished.

Open the MP4 separately and review it at the size your audience will see. A dashboard that looks readable in a large editor window may become illegible in a feed. Check the opening frame, the main action, the result, and the last frame. Then play the entire file with sound, followed by one pass muted to assess whether the visual explanation still works.

For the saved-view example, the final check is specific: can someone identify the filter, read the saved name, and understand that reopening the view restores the intended task list? If any of those moments is unclear, fix that scene before worrying about a more elaborate ending.

Keep the editable project and its source media alongside the approved export. CoAnimator’s project documentation explains project folders and .coa packages. Preserve the assets the project references; an MP4 alone cannot carry its editable structure. Record the approved version so that the next update starts from the right source.

Common problems and how to resolve them

Most problems become easier to diagnose when you identify which part of the workflow failed. A missing command, an unclear scene, and a failed export need different fixes.

Budget for the parts you actually use. Antigravity access, CoAnimator export entitlements, and optional provider-generated audio or imagery are separate considerations. Check current plan details before promising a specific resolution or an unrestricted export. Generating replacement assets can also introduce provider charges beyond the editing workflow itself.

For a first project, reduce the number of moving parts: one short recording, a small set of approved screens, a clear title, and a deliberate result hold. Once that sequence works, you can assess whether voiceover or generated artwork adds enough value to justify the extra production step.

Frequently asked questions

Can Google Antigravity create a video from a prompt?

Yes, with a video creation workflow connected to it. With CoAnimator, the prompt can guide the agent to construct and revise an editable project, then render an MP4. The specificity of the brief and the quality of the source assets still matter. A model response by itself is not an exported video.

Can Antigravity edit an existing MP4?

It can help direct edits through CoAnimator, including selecting footage and changing the surrounding presentation. An imported MP4 remains flattened footage: access to the file does not recover its original text layers, animation controls, or timeline. Keep the source project when you expect to revise those elements later.

Do I need Gemini 4 Argon to use this workflow?

The workflow is built around Antigravity operating a CoAnimator project, rather than a requirement for one named model. Argon is relevant to the current release and Rege’s example, but check the models available to your account. Do not assume a different model will reproduce the same creative result.

Can I make a vertical version from the same project?

You can request a version for a different frame shape, but review its composition separately. Reframe the important action, enlarge text where necessary, and reposition captions. A dashboard that works in a wide frame often needs a closer crop or a different sequence in a narrow one.

Start with a video you can keep improving

Google Antigravity can help create and edit videos when it has the project context, tools, and clear instructions to do the work. CoAnimator provides the editable video workflow and export path; your brief and review determine whether the result explains the product accurately.

Start with one useful action. Give the agent real assets, agree on the sequence, and review the moments where the viewer needs to read or understand something. Then export and watch the actual file. That process gives you more than a first video: it leaves a source project you can revise when the product changes.