To build a small local web app with Google Antigravity, use the IDE, connect a project to a local folder, and have the agent work through a defined task. Then run the app and test it yourself. In a reported Vintage Guitar Vault build, an InfoWorld author started with an Astro page of guitar data and developed it into an app using HTMX and a local SQLite database.

Choose Antigravity IDE

Google distinguishes the standalone Antigravity application, Antigravity IDE, Antigravity CLI, and SDK. The standalone application manages agents; the IDE is the developer-oriented environment for working with code; the CLI provides terminal-based agent interaction. The IDE is the surface used for the local web-app workflow described here. Google’s getting-started guide explains the product distinctions.

Set up the workspace

In a beginner walkthrough, Mikey No Code demonstrates Antigravity IDE setup, plan review, targeted edits, and browser testing with a separate Node.js to-do app.

For its beginner exercise, Google lists macOS, Windows, and selected Linux distributions, along with Chrome and a personal Gmail account. Google describes a project as a scope for one or more folders, with project-specific agent settings and permissions. Connect the folder containing your app before asking the agent to work with its files.

Mikey No Code’s beginner walkthrough demonstrates connecting a local project folder, reviewing a plan, making file-targeted requests, and testing in a browser. Its example is a Node.js to-do app, separate from the Astro guitar-inventory build.

Build the local app

InfoWorld’s author reports using Antigravity IDE to build a local guitar inventory with Astro, HTMX, and SQLite through better-sqlite3. Follow the sequence of that example:

  1. Connect the project to its local folder. Create or open an Antigravity project scoped to the folder where the app will live. A beginner tutorial by Mikey No Code also identifies the project-and-folder setup as the point at which the agent can work with local files.
  2. Create the Astro project. In the IDE terminal, run npm create astro@latest vintage-guitar-vault. The example’s main page is src/pages/index.astro.
  3. Start with guitar data on the page. The reported build began with a hard-coded list. InfoWorld says an Astro frontmatter block, delimited by ---, was needed for the inline completion used in its example to trigger. A comment describing the desired guitar fields also produced more targeted sample data than an unconstrained array prompt.
  4. Add the maintenance request. The example uses HTMX to request /api/maintenance/${guitar.id}. InfoWorld reports that inline assistance supplied markup syntax, while the author used Agent Manager to handle cross-file work, including adding the HTMX script and correcting Astro interpolation.
  5. Move the records into a local database. The author reports using SQLite through better-sqlite3, adding a seed script, and creating the /api/maintenance/[id] endpoint for maintenance history. Those tools and paths belong to this example’s implementation.
  6. Run and test the app. InfoWorld’s author ran npm run dev and viewed the local app at localhost:4321. Test the maintenance feature in the running app as well as checking that the page renders; the example’s reported debugging included a maintenance-feature error.

InfoWorld’s reported Astro, HTMX, and SQLite build shows how the static page grew into a multi-file app. The author reports that a Windows build encountered a node-gyp error, which the agent fixed after approval. Later, a TypeError appeared while testing the maintenance feature; browser testing and further agent work resolved it in that example.

Review changes and test the result

Google documents /plan as a way to request an implementation plan. Possible artifacts include task lists, plans, walkthroughs, diffs, and screenshots. Google cautions that artifacts are not always generated and that an agent may sometimes proceed without a visible Proceed button. Review the changes and relevant permissions when those controls are available, then check the running app in a browser. A proposed plan or generated code is not a substitute for testing the feature you asked the agent to build.