We have been working hard on improving the Antigravity docs and have introduced some new features: - doc search with command bar - copy / view as markdown buttons - better mobile layout - dark mode

Check the docs out today!

@rodydavis Are there any plans to add an integrated terminal to AntiGravity? 🥺 It’s honestly the biggest issue with this harness right now.

That is a great feature request, will pass on the feedback to the team!

💡 Feature Idea: A universal reasoning slider by the prompt bar instead of separate per-model menus! 🎚️ ​Drag across: Low ➔ Med ➔ High ➔ Extra ➔ Max ​Lets users tune thinking budgets per task—supercharging Flash models for coding while keeping the UI fast & clean!

You're DevRel, so here's real dev feedback: we built a rules engine that injects team conventions when the agent touches a matching file. Runs on Claude Code hooks. Antigravity's tool hooks can see the file but can't inject. Who should we talk to?

Adding @NoheDev ! Are you following these docs?

@NoheDev Yes, those exact docs, and they show the gap: PreToolUse output is decision, reason, permissionOverrides. injectSteps only exists on PreInvocation/PostInvocation, which carry no toolCall. The hook that sees the file can't inject. Filed here:

@NoheDev And a companion for the lifecycle side, no session-level event and no agent identity on tool calls made inside subagents: Thanks for pulling in @NoheDev!

@rodydavis For the SessionStart hook event can you give me some context on the use case for this? I am really interested in the resumable case. From my understanding, when we resume a conversation, we reload the context that was previously injected. Are you not seeing this?

@rodydavis (1/4) If resume replays injected steps, the plain resume case is covered. The crux: our injections would ride ephemeralMessage, and the docs define it as a transient system message. Does a transient step survive the reload? If not, injected rules are exactly what resume loses.

@rodydavis (2/4) Beyond resume, two cases drove the SessionStart ask. Re-injection after context loss: on Claude Code it also fires after compaction, where summaries dilute verbatim rules. And once-per-conversation semantics: invocationNum counts model calls, not conversations.

@rodydavis (3/4) To be fair, we can emulate that one with our own state keyed by conversationId, so it's ergonomics, not a blocker like the injection gap. Two questions decide it: do ephemeral steps survive resume, and does agy ever summarize long trajectories?

Mobile layout? Did we finally get a mobile app?

You could install it as a PWA, but this is for the website!

I know I was just hoping we were getting a mobile app, it would make working on the go so much easier.

That is great feedback!

Thanks for putting the work in, I use Antigravity daily.

Thanks for building with us!

Took Antigravity for a test run today, to help me finish off this component.

Ooh, that gives me two additions to my own blog. Really like the layover search and the Markdown buttons. 👏 That shall be done by @antigravity in a blink of an eye. 😏

Im not sure if relevant to write you but 3.7 low is great! I do not need speed but reasoning the model adapt pretty well. It seems im going to launch full app made by antigravity tools within a week. Thanks !

Antigravity was originally designed to let users do multi-model writing, but Claude Sonnet 4.6 and Opus 4.6 are a bit outdated now they may struggle in comparisons with newer models. Gemini's Pro series also hasn't seen any progress,so I hope you can update the third-party models

定位到一个 Antigravity CLI 1.1.13 的疑似回归： 使用的仍是账号 Auth 登录，并非 API Key。但升级后，新会话里 /teamwork-preview 无法识别，而升级前的旧会话仍可正常使用。 怀疑 v1.1.13 新会话内部走了新的 API path，导致 Teamwork 权限检测被误判。

When will we get a context window indicator, and more importantly, the removal of that horrible checkpoint system ? 🙏❤️

Can you elaborate on the issues with the checkpoint system?

The problem is actually quite multi-layered, but the checkpoint system, I think, is really the one that makes things the most annoying ! 🙈 – Context window limit – That is to say, in Antigravity, we have a context window that is already extremely low : 128 000 tokens for the model. For a model that supports one million, it is still a huge shame. – Autocompaction problems – Second thing, and this is why I've responded that in the previous answer: this checkpoint system makes the usable context globally even worse in my opinion 😅 Every 3 to 4 messages, a session summary with a numbered checkpoint is created, this checkpoint making between 2500 and 5000 characters on average. The problem is the following : the context is already low and auto-compactions happen very regularly, but a second problem is added when the summary itself of the previous context is not valid, or at least is way too small, and that moreover it does not retain everything that was done long before. – Loss of debugging context – The problem, if we code in a large professional codebase and with very precise needs, back-end coding or even front-end but complex, the problem being that if a decision after 30 minutes must be reviewed or if we must detect a chain of bugs having started from point A, ending at point C and having continued through point B... Point A is going to be removed, point B is going to be summarized very partially, and point C, the point that leads to the consequence, is the one that will be retained, but without knowing the cause 🤯 It is an example, but it is what creates a whole problem with this context ! – Customization limits – Not to mention that we cannot customize Antigravity much either. Which means that what could lead us to better compensate the lack of context by a better customization, we cannot even do it. In the customization section, we are limited to 20 000 characters. A single skill can consume that ! Our instructions will therefore be eaten behind. The rules, the MCPs, it is the same. These three budgets are put in a single one of 20 000 characters, it is not possible. For the rules, it would need at least 50 000, for the skills at least 250 000, and for MCPs, a number that would be coherent. But currently, it does not work. To give you a little contextual information with the screenshots I will attach : my system instructions stop at only 7 552 characters in the settings (even though the UI says "Token Usage", it actually counts characters !). But the reality is that when I go to Antigravity IDE to code into my VPS (my company's server), I have a real instruction file that Gemini MUST use, and it is more than 102 204 characters long ! ✨ There are so many rules to follow and important operational context in there that it necessarily takes a lot of space for the model to know what to do. Since I cannot put them in the native rules file (because it would be truncated anyway), I am forced to make it read this file at the beginning of each session. The problem is that doing this massively reduces its useful context window right from the start. And moreover, since this content is assimilated as code by the system, it is very quickly forgotten because of the auto-compaction after only three messages 😶 That's another reason why this push to have a very lightweight context for Gemini does way more harm than good 😭❤️ – A great model held back – I cannot find it a satisfying experience to have a model that gets bottlenecked by its own context window, because Gemini 3.7 is far from bad, the model is good ! But its context window breaks everything. I tested it in other environments with much more allowed context, the model is very, very good. But this context window breaks everything. For a coding environment, it is not serious. Grok offers 500 000 tokens, OpenAI now offers more than 256 000 tokens and if we activate it in the settings we can even go up to 400 000 – and Claude is at 1 million. I mean, 128 000 is just not possible anymore, we are in 2026 ;-; – Conclusion – For me, this is where the problem lies. It is a huge problem in the end, because Antigravity really has the power to give developers an immense inspiration for their projects, and I find that from a certain project size, Antigravity becomes unusable because of this problem. I find myself sometimes having to go use other AIs to finish my work because of this problem in fact. And it is a shame, because if I could, I would continue with Gemini. 🙏❤️ And honestly, this is the biggest problem in Antigravity, for me. I also made another long post to summarize a few other related issues, if you want to take a look ! :) (Vikas and Mohan had already seen it, but I think that the more Antigravity employees see it, the more you'll be able to understand the problems in it 😄✨)

Still having this issue, bro. It occurs in both 2.55 and 2.52. The model can't read the images and files when I paste the attachment files. I get the issues below.

What is this 😳 project for. Antigravity ?

@antigravity docs and website!

Please bring Google Antigravity to VS Code as an official extension. Many developers already rely on VS Code and don't want to switch to a separate IDE just to use Antigravity. Keep the editor, add the Agent.

Bros spent one month to implement dark mode. Lol

hi rody, i still can’t preview PDF files in antigravity app sidebar, it just displays as unrendered texts rather than normal PDF , please fix this thx

Please add a file navigation function and click to jump to the file directory.

为什么不能在对话框上传 文档

For eg. When antigravity builds a file or whatever, and it says here is your file. Then you click on the file, it just opens the code of the file in that side bar, or tries to. It would be amazing for ease of use if there was a option on that where you opens its explore location

clean docs make a huge difference when building with agent platforms dark mode and markdown export are solid additions for antigravity

Why do you still not allow file uploads from my Mac? Only images

You can drag and drop them for multiple expanded file types, but the ui picker needs to be updated to support the non-image file types too. Will pass on the feedback!

Ok thanks

Well, why don't you use antigravity to speed it up?

We did!

Thanks for improving the website, could you make an antigravity version for Android?

Any plans to improve how file reading and contexts? I noticed too many file reads and re-reading, like same file, same line numbers but the second time happened after it read another files Also any plans to improve how AG edit files? Would like to know

Is it for Antigravity or Antigravity IDE?

This is for the website and docs!

Hmm..is it a changelog for docs😆

This is the release you want to check out!

In AntiGravity 2.0 pls disable inapp notification banners. They block the submit button whenever the agent requests permission, forcing users to manually dismiss the banner first. Push notifications outside the app are fine, but inapp banners disrupt the workflow. (windows 11)

I'm missing the possibility to check in on the progress in antigravity.

Can we add token dashboards and context usage displays? I really do need to say thank you.

Can you please improve on agents to not give stupid errors. Ah and to not struggle with bash si hard

Can you elaborate or share examples?

next time it happens I will directly send you the screenshots so that you know I’m not trolling

Feedback is appreciated and want to track this down!

Have you ever considered making [high medium low] a separate option? At present, only five files can be uploaded?

For what are you referring to?

The model level in antigravity 2.0 and the maximum number of uploaded files in the dialog box.

Looking great. Would love if @antigravity kept input focus after permission requests. Tiny thing, but super annoying mid-typing.

@antigravity Thanks for the feedback!

The website I built off of this thing is crazy I can't wait to make it go live