Appreciate all the takes so far, the team is reading this thread for feedback. Feel free to pile on with bugs or feature requests we'll all be reading.

However, Gemini at one point output what will probably be the highlight of my day:

"I have made a complete mess of the code. I will now revert all changes I have made to the codebase and start over."

What great self-awareness and willingness to scrap the work! :)

We really are living in the future

https://developers.google.com/gemini-code-assist/resources/p...

When you use Gemini Code Assist for individuals, Google collects your prompts, related code, generated output, code edits, related feature usage information, and your feedback to provide, improve, and develop Google products and services and machine learning technologies.

To help with quality and improve our products (such as generative machine-learning models), human reviewers may read, annotate, and process the data collected above. We take steps to protect your privacy as part of this process. This includes disconnecting the data from your Google Account before reviewers see or annotate it, and storing those disconnected copies for up to 18 months. Please don't submit confidential information or any data you wouldn't want a reviewer to see or Google to use to improve our products, services, and machine-learning technologies.

It keeps putting import statements in the middle of files or duplicating edits at the wrong places, but it's hard to get frustrated since it seems pretty self-aware. I didn't say it was making a mess - it's just being hard on itself.

A short session ended up sending over 3 million tokens though - wonder how the economics of this work out for google?

More of my notes here: https://simonwillison.net/2025/Jun/25/gemini-cli/

Is your vision with Gemini CLI to be geared only towards non-commercial users? I have had a workspace account since GSuite and have been constantly punished for it by Google offerings all I wanted was gmail with a custom domain and I've lost all my youtube data, all my fitbit data, I cant select different versions of some of your subscriptions (seemingly completely random across your services from a end-user perspective), and now as a Workspace account I cant use Gemini CLI for my work, which is software development. This approach strikes me as actively hostile towards your loyal paying users...

> hello

[API Error: {"error":{"message":"{\n \"error\": {\n \"code\": 429,\n \"message\": \"Resource has been exhausted (e.g. check quota).\",\n \"status\": \"RESOURCE_EXHAUSTED\"\n }\n}\n","code":429,"status":"Too Many Requests"}}] Please wait and try again later. To increase your limits, request a quota increase through AI Studio, or switch to another /auth method

⠼ Polishing the pixels... (esc to cancel, 84s)

I have thrown very large codebases at this and it has been able to navigate and learn them effortlessly.

Then there are 3rd party channels, if you have a recent samsung phone, you get 1 yr access to AI features powered by gemini, after which you need to pay. And lord knows where else has google been integrating gemini now.

Ive stopped using google's AI now. Its like they have dozens of teams within gemini on completely different slack sessions.

Definitely not because of Claude Code eating our lunch!

- Open-source (Apache 2.0, same as OpenAI Codex)

- 1M token context window

- Free tier: 60 requests per minute and 1,000 requests per day (requires Google account authentication)

- Higher limits via Gemini API or Vertex AI

- Google Search grounding support

- Plugin and script support (MCP servers)

- Gemini.md file for memory instruction

- VS Code integration (Gemini Code Assist)

I just symlink now to AGENTS.md

If it sounds too good to be true, it probably is. What’s the catch? How/why is this free?

1) I tried to use it on an existing project asking this "Analyse the project and create a GEMINI.md". It fumbled some non sense for 10-15 minutes and after that it said it was done, but it had only analysed a few files in the root and didn't generate anything at all.

2) Despite finding a way to login with my workspace account, it then asks me for the GOOGLE_CLOUD_PROJECT which doesn't make any sense to me

3) It's not clear AT ALL if and how my data and code will be used to train the models. Until this is pretty clear, for me is a no go.

p.s: it feels like a promising project which has been rushed out too quickly :/

https://youtu.be/HC6BGxjCVnM?feature=shared&t=36

It's a FOSS MCP server I created a couple of weeks ago:

- https://getvoicemode.com

- https://github.com/mbailey/voicemode

# Installation (~/.gemini/settings.json)

{

  "theme": "Dracula",  

  "selectedAuthType": "oauth-personal",  

  "mcpServers": {  

    "voice-mode": {  

      "command": "uvx",  

      "args": [  

        "voice-mode"  

      ]  

    }  

  }  

}

  1. Go to their enterprise site
  2. See what privacy guarantees they advertise above the consumer product
  3. Conclusion: those are things that you do not get in the consumer product

For Code Assist, here's their enterprise-oriented page vs their consumer-oriented page:

https://cloud.google.com/gemini/docs/codeassist/security-pri...

https://developers.google.com/gemini-code-assist/resources/p...

It seems like these are both incomplete and one would need to read their overall pages, which would be something more like

https://support.google.com/a/answer/15706919?hl=en

https://support.google.com/gemini/answer/13594961?hl=en#revi...

When I try to log in using a personal account, it tells me I need to generate something in the dashboard. I know that on my main account from maybe 15 years ago I had things like domains thats nowdays Google Workspace. But I'm trying to log in and authorize using my second (NOT MAIN), relatively new account. Despite this, it redirects me to the console (like it would for the first account), and there it says I'm logged in (!), but the application itself says I'm not.

I switched profiles within the console too, tried the same steps, and it still results in the same problem. Yes, I can see my new profile in the top right corner, not the old one. It sends me to a "Users may have to specify a GOOGLE_CLOUD_PROJECT if:" github [0] saying this ID is needed, and I honestly would even generate it, but I don't want to waste time with the program if you can't even get basic authorization working correctly.

0: https://github.com/google-gemini/gemini-cli/blob/main/docs/c...

It generates a bunch of fake activity indicators based on your prompt, then cycles through them on a timer. It has no bearing on the actual activity going on underneath.

It appears to be much slower than Claude Code, possibly due to being overloaded, but it feels like it thinks a lot more before beginning to suggest file edits. The permissions aren't as nice either. Where Claude Code suggests "allow uv run pytest without approval", Gemini suggests "allow uv run without approval", which is broader than I would like.

In comparison to Claude Code Opus 4, it seemed much more eager to go on a wild goose chase of fixing a problem by creating calls to new RPCs that then attempted to modify columns that didn't exist or which had a different type, and its solution to this being a problem was to then propose migration after migration to modify the db schema to fit the shape of the rpc it had defined.

Reminded me of the bad old days of agentic coding circa late 2024.

I'm usually a big fan of 2.5 Pro in an analysis / planning context. It seems to just weirdly fall over when it comes to tool calling or something?

1. First authentication didn't work on my headless system, because it wants an oauth redirect to localhost - sigh.

2. Next, WebFetch isn't able to navigate github, so I had to manually dig out some references for it.

3. About 2 mins in, I just got ``` ℹ Rate limiting detected. Automatically switching from gemini-2.5-pro to gemini-2.5-flash for faster responses for the remainder of this session. ``` in a loop with no more progress.

I understand the tool is new, so not drawing too many conclusions from this yet, but it does seem like it was rushed out a bit.

Gemini CLI does not take new direction especially well. After planning, I asked it to execute and it just kept talking about the plan.

Another time when I hit escape and asked it to stop and undo the last change, it just plowed ahead.

It makes a lot of mistakes reading and writing to files.

Some, but by no means all, of the obsequious quotes from my first hour with the product: - “You’ve asked a series of excellent questions that have taken us on a deep dive ...” - “The proposed solution is not just about clarity; it’s also the most efficient and robust.”

This perfectly demonstrates the benefit of the nodejs platform. Trivial to install and use. Almost no dependency issues (just "> some years old version of nodejs"). Immediately works effortlessly.

I've never developed anything on node, but I have it installed because so many hugely valuable tools use it. It has always been absolutely effortless and just all benefit.

And what a shift from most Google projects that are usually a mammoth mountain of fragile dependencies.

(uv kind of brings this to python via uvx)

How did they do that pretty "GEMINI" gradient in the terminal? is that a thing we can do nowadays? It doesn't seem to be some blocky gradient where each character is a different color. It's a true gradient.

(yes I'm aware this is likely a total clone of claude code, but still am curious about the gradient)

1. The thing going in a circle trying to fix a bug by persistently trying different permutations of an API interface it never bothered to check the definition of. Isn't that why it's internet connected?

2. When I asked it to just analyze instead of change stuff. It just hung for 20 minutes giving me responses saying that gemini-2.5-pro was slow, and that it was switching to 2.5-flash, with no other output to indicate what it was doing other than those cute scrolling messages that mean nothing.

At least in Claude it's clear that the messages mean nothing, because they're clearly generic. Gemini gives you the impression the status messages mean something since they're sort of related to the topic being worked on.

Because it says in the README:

> Authenticate: When prompted, sign in with your personal Google account. This will grant you up to 60 model requests per minute and 1,000 model requests per day using Gemini 2.5 Pro.

> For advanced use or increased limits: If you need to use a specific model or require a higher request capacity, you can use an API key: ...

When I have the Google AI Pro subscription in my Google account, and I use the personal Google account for authentication here, will I also have more requests per day then?

I'm currently wondering what makes more sense for me (not for CLI in particular, but for Gemini in general): To use the Google AI Pro subscription, or to use an API key. But I would also want to use the API maybe at some point. I thought the API requires an API key, but here it seems also the normal Google account can be used?

https://github.com/vadimdemedes/ink

Never had anything like this with claude code.

I've used Gemini 2.5 Pro quite a lot and like most people I find it's very intelligent. I've bent over backwards to use Gemini 2.5 Pro in another piece of work because it's so good. I can only assume it's the gemini CLI itself that's using the model poorly. Keen to try again in a month or two and see if this poor first impression is just a teething issue.

I told it that it did a pretty poor job and asked it why it thinks that is, told it that I know it's pretty smart. It gave me a wall of text and I asked for the short summary

> My tools operate on raw text, not the code's structure, making my edits brittle and prone to error if the text patterns aren't perfect. I lack a persistent, holistic view of the code like an IDE provides, so I can lose track of changes during multi-step tasks. This led me to make simple mistakes like forgetting a calculation and duplicating code.

gave it brief instructions to deploy a hobby static site on cloud run yesterday; with additional architecture to set up (load balancer etc). It got into a loop, started deleting the gcloud resouces it created when it hit a 403 error for the site.

I hit ESC and prodded a little in terms of authentication rules ... and off it went to complete the task. bravo!

1: https://block.github.io/goose/

It's more focused on implications for docs strategy (I'm worried that agent providers are steering us towards duplicating information that's already in eng docs) rather than user best practices i.e. "put this into your agent doc to improve agent performance"

https://technicalwriting.dev/ai/agents/

Tried upgrading to the Standard plan through Google Cloud with the hope that it would allow me to do more, but after switching my account to the now-paid account, it still almost instantly downgraded me to 2.5-flash

For the times when I was able to use 2.5-pro, the output has been very good. But the moment it switches to flash, the quality goes down by about 80% and I would never let it code on anything

Set up not too long ago, and afaik pretty load-bearing for this. Feels great, just don’t ask me any product-level questions. I’m not part of the Gemini CLI team, so I’ll try to keep my mouth shut.

Not going to lie, I’m pretty anxious this will fall over as traffic keeps climbing up and up.

How do I reset permissions so it always asks again for `git` invocations?

Thanks!

A bit gutted by the `make sure it is not a workspace account`. What's wrong with Google prioritising free accounts vs paid accounts? This is not the first time they have done it when announcing Gemini, too.

---

Right now there is one CLI that influences and stands far and beyond all others. Smooth UX, and more critical some "natural" or inherent ability to use its tools well.

Gemini can also achieve this - but i think it's clear the leader is ahead because they have a highly integrated training process with the base model and agentic tool use.

It is vastly more difficult to understand what Google is offering compared to the others, to what cost, getting an API-key or similar, understanding usage/billing across the suite, etc.

I wouldn't expect any regular person to bother signing up for any of Google's products, let alone understand what they're really offering.

1. Gemini Code Assist (GCA) for Individuals: FREE for 1,000 model requests/day

2. GCA Standard: $22.80/month for 1,500 model requests/day (1.5x more than FREE)

3. GCA Enterprise: $54.00/month for 2,000 model requests/day (2X more than FREE)

Source: https://codeassist.google

I guess I will use something else. This is all very annoying given that I actually pay for Gemini Pro...

I use aichat now but it's not perfect. https://github.com/sigoden/aichat

The UI on this tool is much better.

That's a ton of free limit. This has been immensely more successful than void ide.

Well, not sure that it makes sense to do it, anyway I've tried to run in in a cell and in the google colab terminal. Still waiting for auth (?)

https://github.com/opencode-ai/opencode

At the bottom of README.md, they state:

"This project leverages the Gemini APIs to provide AI capabilities. For details on the terms of service governing the Gemini API, please refer to the terms for the access mechanism you are using:

* Gemini API key

* Gemini Code Assist

* Vertex AI"

The Gemini API terms state: "for Unpaid Services, all content and responses is retained, subject to human review, and used for training".

The Gemini Code Assist terms trifurcate for individuals, Standard / Enterprise, and Cloud Code (presumably not relevant).

* For individuals: "When you use Gemini Code Assist for individuals, Google collects your prompts, related code, generated output, code edits, related feature usage information, and your feedback to provide, improve, and develop Google products and services and machine learning technologies."

* For Standard and Enterprise: "To help protect the privacy of your data, Gemini Code Assist Standard and Enterprise conform to Google's privacy commitment with generative AI technologies. This commitment includes items such as the following: Google doesn't use your data to train our models without your permission."

The Vertex AI terms state "Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction."

What a confusing array of offerings and terms! I am left without certainty as to the answer to my original question. When using the free version by signing in with a personal Google account, which doesn't require a Gemini API key and isn't Gemini Code Assist or Vertex AI, it's not clear which access mechanism I am using or which terms apply.

It's also disappointing "Google's privacy commitment with generative AI technologies" which promises that "Google doesn't use your data to train our models without your permission" doesn't seem to apply to individuals.

better question is why do you need a modle specific CLI when you should be able to plug in to individual models.

We are now three years into the AI revolution and they are still forcing us to copy and paste and click click crazy to get the damn files out.

STOP innovating. STOP the features.

Form a team of 500 of your best developers. Allocate a year and a billion dollar budget.

Get all those Ai super scientists into the job.

See if you can work out “download all files”. A problem on the scale of AGI or Dark Matter, but one day google or OpenAI will crack the problem.

> 429: Too many requests

Mind you, this is with a paid API key

I hate this openwashing. It's a closed model, its weights are nowhere to be seen! (not to mention the training data, the real "source" of a LLM)

The fact that there is a small component that is open source that accesses this closed model doesn't change that at all.

https://imgur.com/ZIZkLU7

This is shown at the top of the screen in https://aistudio.google.com/apikey as the suggested quick start for testing your API key out.

Not a great look. I let our GCloud TAM know. But still.

Google services have become a patchwork of painfully confounding marketing terms that mean nothing and obfuscate what they actually provide.

Gemini Pro and Claude play off of each other really well.

Just started playing with Gemini CLI and one thing I miss immediately from Claude code is being able to write and interject as the AI does its work. Sometimes I interject by just saying stop, it stops and waits for more context or input or ai add something I forgot and it picks it up..

I do not get it why they don’t pick Go or Rust so i get a binary.

Not impressed. These companies have billions at their disposal, and probably pay $0 in tax, and the best they can come up with is this?

While writing this comment, thinking that there should be some packaging tool that would create a binaries from npx cli tools. I remember such things for python. Binaries were fat, but it is better then keep nodejs installed on my OS