Google Antigravity export: What survives one-click handoff to local AI apps

Google’s AI Studio is a powerful sandbox for prototyping multi-agent apps, but every developer eventually faces the same frustration: a working prototype locked inside a browser tab. At I/O 2026, Google introduced Export to Antigravity, a one-click feature designed to bridge that gap by moving prototypes to a local production workspace while preserving context. To separate hype from reality, we put this export tool through its paces with a fully functional two-agent research project. Here’s what we found.

A real prototype meets the export button

Our test case was Research Digest, a sequential two-agent system built entirely in Google AI Studio’s Build mode. Agent 1, the Researcher, used grounded web search to collect sources on a given topic, while Agent 2, the Editor, synthesized those findings into a polished digest. The project relied on Firestore for persistence and maintained an archive of past digests.

Before even attempting the export, we encountered our first obstacle—a setup warning during Firestore provisioning in AI Studio. While not directly related to the export itself, it highlighted a recurring theme: even seemingly small configuration steps can derail workflows before migration even begins.

The export process itself was surprisingly straightforward. From the Code tab, we selected Export → Export to Antigravity. The dialog provided clear, upfront transparency: all project files, conversation history, and a single secret would be included. The absence of jargon or hidden steps made the experience feel intentional and user-friendly.

What actually made the journey intact

Google’s export dialog made several explicit promises about what would transfer. We verified each claim carefully.

• All project files – The full folder structure arrived intact, including /.agents, /.antigravity, /src, configuration files, and a detailed README.md with setup instructions. No files were missing or corrupted.

• One secret – The GEMINI_API_KEY was correctly populated in the .env file and worked immediately after export. No manual re-entry was required, eliminating a common migration pain point.

• Conversation history – Despite the dialog’s explicit claim, this did not transfer. The imported project in Antigravity’s Agent Manager displayed “No conversations yet,” even though the AI Studio workspace contained extensive prior exchanges. We confirmed this result twice across different screens.

The discrepancy between the export dialog’s promise and the actual outcome underlines the importance of managing expectations when moving between cloud and local environments.

The hidden obstacles that slowed progress

Three unadvertised challenges emerged during migration, none directly related to the export’s core functionality but critical to the developer experience.

1. Location, location, location

After clicking Export, the dialog closed without indicating where the project files were saved. A search of the Downloads folder yielded nothing. Only by searching the entire system by project name did we locate the project—nested deep inside ~/antigravity/.

The Antigravity ecosystem operates through two distinct interfaces: the Agent Manager (the chat and orchestration surface) and the Antigravity IDE (a VS Code-based editor). These tools do not communicate automatically, so even discovering the project’s physical location required manual effort. Once found, we could finally use the “Open Folder” option within the IDE to view the code.

2. Conversation context vanished without notice

While the export dialog claimed conversation history would transfer, the reality was starkly different. The imported project showed no prior interactions in the Agent Manager, meaning any contextual threads from AI Studio were lost. This missing context can disrupt debugging and iteration, especially when troubleshooting complex agent behaviors.

3. A pre-existing bug surfaced after migration

After running npm install and npm run dev with zero dependency vulnerabilities, the app connected successfully to Cloud Firestore. However, attempting to generate a digest triggered an error:

Tool use with a response mime type: 'application/json' is unsupported

This error stemmed from a constraint in the Gemini API: combining tool use (such as web search) with forced JSON-mode output in the same call is not supported. Critically, this bug was already embedded in the generated code—it was not introduced by the export process itself.

Rather than debugging manually, we leveraged Antigravity’s built-in agent debugging system. By pasting the error into the agent panel, the system analyzed four source files—gemini.ts, server.ts, App.tsx, and DigestViewer.tsx—and proposed a targeted fix. The solution involved adjusting Agent 1 to return plain text instead of forced JSON, allowing Agent 2 to parse the content into a structured digest. The diff showed 59 lines added and 53 removed across two files. After applying the changes and rerunning the process, the digest generation worked end to end, with sources properly cited and persisted to Firestore.

A practical pre- and post-export checklist

Moving from cloud prototyping to local development requires more than a single click. Here’s a concise checklist to avoid common pitfalls.

Before exporting:

• Confirm the list of secrets attached to your AI Studio project matches your expectations—the export UI will explicitly name each one.
• Do not assume conversation history will transfer; copy any critical context manually beforehand.

After exporting, before resuming development:

• Locate the project folder yourself by searching for the project name or checking ~/antigravity/—the UI does not always surface this information.
• Open the Antigravity IDE separately from the Agent Manager to view and run the code.
• Run the application end to end at least once before assuming it works—the export may succeed while underlying code still contains bugs.
• If a runtime error occurs, describe it directly to Antigravity’s agent debugging panel before diving into manual inspection. The agent can often diagnose and patch issues across multiple files in a single pass.

The export process delivered on its core promises: files and secrets migrated cleanly, and Firestore connectivity—previously broken in AI Studio—functioned correctly in the local environment with no additional configuration. What didn’t survive was conversation context, and what consumed the most time wasn’t the migration itself but discovering where the project landed and addressing a pre-existing code generation bug.

Antigravity’s agent-assisted debugging fixed that bug faster than manual intervention would have allowed. In total, the one-click export took moments, but the real work required about fifteen minutes of focused troubleshooting. If you’re considering migrating AI Studio prototypes to Antigravity, move forward with clear expectations—and a willingness to double-check the details.