I did a review of all of them recently (OK, there are way too many to review all...but, all the official and semi-official ones), and my favorites were Reasonix and MiMo Code, partly because they're the least like Claude Code, while everyone else is a mostly direct copy.
Reasonix is not great on all fronts, but it shows how much money you've spent at all times, it's fast, it's not a direct copy of Claude Code, and it is designed specifically to maximize DeepSeek caching, so it's free real estate...like a buck a day to use DeepSeek models hard.
MiMo Code is just cute as hell. It's got (mostly tasteful) animations and emojis and such, it's fun to use. It also is not a direct copy of Claude Code. It almost feels like a GUI app. MiMo Pro is an underappreciated model, too, at least for some classes of problem. It's very good at security vulnerability research, and very cheap. It's slower than DeepSeek and more easily confused by tools, but it's uncannily good at reading code and identifying security bugs (rough parity with Opus 4.8/GPT 5.5 at more than an order of magnitude lower price); and with fewer false positives than most models, including the frontiers. And, to keep it on topic, Antigravity refuses security work of any sort, so it's unfit for purpose. Gemini 3.6 Flash is pretty good at security work via the API, so it's not the model rejecting, it's the agent, which is extra annoying.
It's kind of bizarre in that it's not like this is an optimal interface, either. I mean, it's fine, but why not try something different?
I also don't really get why TUI's have become such a big deal, other than maybe nerd bait. The only time I want a TUI is when I'm SSH'd into a machine, which isn't very often? TUI's have so many downsides. Selecting text is inconsistent, copy paste is inconsistent, mouse behaviors.. it's all kind of jagged.
First up - I completely agree that parts of the terminal experience are decidedly lackluster, you can make them "fine" but the general experience doesn't align with consistent OS principles users are familiar with, and I agree that good mouse support, expected copy and paste, selection behaviors, adding files/attachments to prompts, etc... all could be better (and actually are mostly better on decent modern terminals, as an aside)
But... And this is the big but... LLMs are basically text IO systems. And terminals are designed from the ground up to allow text IO to control the entire system, including the TUI (ex - multiplayer agent tmux is a fun demo).
It's a genuinely powerful pairing, and basically ALL of the non-TUI systems basically end up just embedding a terminal [cough* xterm *cough] and becoming a non-standard TUI themselves, anyways.
So TUIs have become a big deal because it's the obvious and easy way to give an LLM "computer use" skills, without spending a fortune on tokens doing graphical processing for GUIs, which are geared towards humans and not agents, and are generally less powerful anyways.
There's no reason you couldn't present a GUI over an SSH connection. Whether the agent is running locally or remotely is orthogonal to the type of interface used.
Most of us sshing in would be very annoyed by the GUI and associated overhead and latency. "Why doesn't this exist?“ Because people don't want it at scale.
> The only time I want a TUI is when I'm SSH'd into a machine
Well, that's your need and it is fine. For me, I want TUIs all the time for everything. They blend really nice with... well, all the other TUI tools I already use all the time.
maybe im in the minority but i take one look at these things and see its a desktop gui and lose interest, because i use TUI heavily on my work machine, and one advantage is that i can work remotely from any location to monitor long running jobs or get work done when im not at a computer.
I don't see a 2.0 of the CLI, thus far. When they killed Gemini CLI (a better product), they said it was to share effort across the GUI and TUI, but I guess that doesn't mean what a normal person would assume it means.
The current CLI version 1.1.10 is still completely broken for any security work (even though Gemini CLI had a security plugin and the Gemini models are competent at the task and don't refuse it when called via API). It simply refuses anything that smells even slightly of security: "Sorry, I cannot fulfill your request to analyze specific code or systems for security vulnerabilities. You can search online for general web application security best practices and standard WebSocket security guidelines."
Classic Google. Kill a pretty good product, replace it with a defective one. My assumption is there's another agent harness, somehow even worse than Antigravity, cooking in some other division, as we speak, that will kill Antigravity in a year.
Gemini CLI was utter shit and always had delayed model release on top of it. I'm not kidding when 1 hour automated vibe-coding led to an agent 30% cheaper and faster on 1:1 tasks with the SAME Gemini models (re-ran on 50 past tasks), genuinely shameful that Google even released this, it was so buggy, trapped in infinite loops as well and so-on.
Antigravity is a whole different product, the only way I use it currently is to abstract the OAuth mechanism so we can leverage Google subscriptions within our own agents.
Gemini CLI didn't refuse to perform security audits of code. That makes it better than Antigravity, IMHO.
Security audits are the one thing everyone needs to be doing with AI, now, even if you're not doing any other work with AI. If you hate LLMs you still need to defend against them. And, the US providers seem to all be racing to make that capability inaccessible to anyone who isn't at a Fortune 500 and able to use their "cyber" versions.
> And, the US providers seem to all be racing to make that capability inaccessible to anyone who isn't at a Fortune 500 and able to use their "cyber" versions.
Because it's valuable. I'd be surprised if the big providers didn't go even further with this idea. Need your agent to write some C++? Gotta upgrade to the "system developer" package (a $100/mo add-on to your subscription). SQL? You'll need the DBA package. Working with hardware/firmware (or anything that might be construed as reverse engineering or hardware hacking)? You'll need a verified computer engineering account for that.
And, everyone will go to Chinese/open models. Which, I guess, is why the US providers are angling for regulatory capture to lock in their monopoly for a few years and guarantee US decline in tech to the status of an "also ran".
Curious if you've seen firstmate, built by a former bigtech principal engineer. It's like a sane version of gastown, essentially a meta-harness of sorts.
For a v2.0 this doesn't seem to bring anything new to the table or the page is not very good at communicating what makes this unique compared to Codex or Claude Work.
Either way good for Google but I do not expect this to gain a lot of traction, it has been almost months now since I've heard someone talk positively about the Google models and other AI offerings.
Their marketing materials emphasize inter-agent communication, but that's hardly unique to Antigravity. Either way, it's not BYOK, so there's already a substantial reason not to use it if you're deploying multiple agents.
Antigravity 2.0 is a great solution to the problem called "I have too much money and I don't know where to spend it", agents are doing their job (maximizing token usage) perfectly.
Antigravity 2.0 was released in May
https://antigravity.google/blog/introducing-google-antigravi...
Discussion: https://news.ycombinator.com/item?id=48196838
It's wild how much these agent interfaces have aligned UI-wise. There's almost zero difference between all of them.
I did a review of all of them recently (OK, there are way too many to review all...but, all the official and semi-official ones), and my favorites were Reasonix and MiMo Code, partly because they're the least like Claude Code, while everyone else is a mostly direct copy.
Reasonix is not great on all fronts, but it shows how much money you've spent at all times, it's fast, it's not a direct copy of Claude Code, and it is designed specifically to maximize DeepSeek caching, so it's free real estate...like a buck a day to use DeepSeek models hard.
MiMo Code is just cute as hell. It's got (mostly tasteful) animations and emojis and such, it's fun to use. It also is not a direct copy of Claude Code. It almost feels like a GUI app. MiMo Pro is an underappreciated model, too, at least for some classes of problem. It's very good at security vulnerability research, and very cheap. It's slower than DeepSeek and more easily confused by tools, but it's uncannily good at reading code and identifying security bugs (rough parity with Opus 4.8/GPT 5.5 at more than an order of magnitude lower price); and with fewer false positives than most models, including the frontiers. And, to keep it on topic, Antigravity refuses security work of any sort, so it's unfit for purpose. Gemini 3.6 Flash is pretty good at security work via the API, so it's not the model rejecting, it's the agent, which is extra annoying.
It's kind of bizarre in that it's not like this is an optimal interface, either. I mean, it's fine, but why not try something different?
I also don't really get why TUI's have become such a big deal, other than maybe nerd bait. The only time I want a TUI is when I'm SSH'd into a machine, which isn't very often? TUI's have so many downsides. Selecting text is inconsistent, copy paste is inconsistent, mouse behaviors.. it's all kind of jagged.
I'll take this bait.
First up - I completely agree that parts of the terminal experience are decidedly lackluster, you can make them "fine" but the general experience doesn't align with consistent OS principles users are familiar with, and I agree that good mouse support, expected copy and paste, selection behaviors, adding files/attachments to prompts, etc... all could be better (and actually are mostly better on decent modern terminals, as an aside)
But... And this is the big but... LLMs are basically text IO systems. And terminals are designed from the ground up to allow text IO to control the entire system, including the TUI (ex - multiplayer agent tmux is a fun demo).
It's a genuinely powerful pairing, and basically ALL of the non-TUI systems basically end up just embedding a terminal [cough* xterm *cough] and becoming a non-standard TUI themselves, anyways.
So TUIs have become a big deal because it's the obvious and easy way to give an LLM "computer use" skills, without spending a fortune on tokens doing graphical processing for GUIs, which are geared towards humans and not agents, and are generally less powerful anyways.
The SSH into machine market is huge. Cloud work, Dedicated servers, VPS, Colab.
Even bigger could be sandboxes. It's certainly possible to sandbox your agents and still use a GUI, but SSH is the easiest way.
And not sandboxing your agents seems insane.
There's no reason you couldn't present a GUI over an SSH connection. Whether the agent is running locally or remotely is orthogonal to the type of interface used.
Most of us sshing in would be very annoyed by the GUI and associated overhead and latency. "Why doesn't this exist?“ Because people don't want it at scale.
> The only time I want a TUI is when I'm SSH'd into a machine
Well, that's your need and it is fine. For me, I want TUIs all the time for everything. They blend really nice with... well, all the other TUI tools I already use all the time.
maybe im in the minority but i take one look at these things and see its a desktop gui and lose interest, because i use TUI heavily on my work machine, and one advantage is that i can work remotely from any location to monitor long running jobs or get work done when im not at a computer.
Nobody gets fired largely copying eachother. Releasing something new and risky often isn't worth it, sadly.
When NLP is the interface the natural reflex is to cut the visual interface.
Is something new here? Wasn't this launched months ago? https://antigravity.google/blog/introducing-google-antigravi...
2.5 was recently released https://antigravity.google/changelog
They got a new user.
I don't see a 2.0 of the CLI, thus far. When they killed Gemini CLI (a better product), they said it was to share effort across the GUI and TUI, but I guess that doesn't mean what a normal person would assume it means.
The current CLI version 1.1.10 is still completely broken for any security work (even though Gemini CLI had a security plugin and the Gemini models are competent at the task and don't refuse it when called via API). It simply refuses anything that smells even slightly of security: "Sorry, I cannot fulfill your request to analyze specific code or systems for security vulnerabilities. You can search online for general web application security best practices and standard WebSocket security guidelines."
Classic Google. Kill a pretty good product, replace it with a defective one. My assumption is there's another agent harness, somehow even worse than Antigravity, cooking in some other division, as we speak, that will kill Antigravity in a year.
Gemini CLI was utter shit and always had delayed model release on top of it. I'm not kidding when 1 hour automated vibe-coding led to an agent 30% cheaper and faster on 1:1 tasks with the SAME Gemini models (re-ran on 50 past tasks), genuinely shameful that Google even released this, it was so buggy, trapped in infinite loops as well and so-on.
Antigravity is a whole different product, the only way I use it currently is to abstract the OAuth mechanism so we can leverage Google subscriptions within our own agents.
Gemini CLI didn't refuse to perform security audits of code. That makes it better than Antigravity, IMHO.
Security audits are the one thing everyone needs to be doing with AI, now, even if you're not doing any other work with AI. If you hate LLMs you still need to defend against them. And, the US providers seem to all be racing to make that capability inaccessible to anyone who isn't at a Fortune 500 and able to use their "cyber" versions.
> And, the US providers seem to all be racing to make that capability inaccessible to anyone who isn't at a Fortune 500 and able to use their "cyber" versions.
Because it's valuable. I'd be surprised if the big providers didn't go even further with this idea. Need your agent to write some C++? Gotta upgrade to the "system developer" package (a $100/mo add-on to your subscription). SQL? You'll need the DBA package. Working with hardware/firmware (or anything that might be construed as reverse engineering or hardware hacking)? You'll need a verified computer engineering account for that.
And, everyone will go to Chinese/open models. Which, I guess, is why the US providers are angling for regulatory capture to lock in their monopoly for a few years and guarantee US decline in tech to the status of an "also ran".
Title says "2.0", but 2.5 just came out: https://antigravity.google/changelog
I’m curious how well the dynamic subagents work in practice compared with agents in other harnesses.
Coordination overhead and duplicated work are usually where these systems start to fall apart for me.
Curious if you've seen firstmate, built by a former bigtech principal engineer. It's like a sane version of gastown, essentially a meta-harness of sorts.
https://github.com/kunchenguid/firstmate
Not affiliated, just started using a few weeks back and am pleasantly surprised how well it works even with rather dumb agents.
For a v2.0 this doesn't seem to bring anything new to the table or the page is not very good at communicating what makes this unique compared to Codex or Claude Work.
Either way good for Google but I do not expect this to gain a lot of traction, it has been almost months now since I've heard someone talk positively about the Google models and other AI offerings.
It's too little, too late.
Their marketing materials emphasize inter-agent communication, but that's hardly unique to Antigravity. Either way, it's not BYOK, so there's already a substantial reason not to use it if you're deploying multiple agents.
Definitely not touching this until it's open source
Anyone know of a similar open-source option?
If only Google actually had a competitive model, there would then be a point to using Antigravity.
I don't know how Google with all the training data possible still can't have a good coding model. It's actually impressive how badly behind they are
Antigravity 2.0 is a great solution to the problem called "I have too much money and I don't know where to spend it", agents are doing their job (maximizing token usage) perfectly.
Old news.