Gemini CLI (opens in new tab)

Thank you for your work on this. I spent the afternoon yesterday trying to convert an algorithm written in ruby (which I do not know) to vanilla JavaScript. It was a comedy of failing nonsense as I tried to get gpt-4.1 to help, and it just led me down pointless rabbit holes. I installed Gemini CLI out of curiosity, pointed it at the Ruby project, and it did the conversion from a single request, total time from "think I'll try this" to it working was 5 minutes. Impressed.

There is one feature in Claude Code which is often overlooked and I haven't seen it in any of the other agentic tools: There is a tool called "sub-agent", which creates a fresh context windows in which the model can independently work on a clearly defined sub-task. This effectively turns Claude Code from a single-agent model to a hierarchical multi-agent model (I am not sure if the hierarchy goes to depths >2).

I wonder if it is a concious decision not to include this (I imagine it opens a lot of possibilities of going crazy, but it also seems to be the source of a great amount of Claud Code's power). I would very much like to play with this if it appears in gemini-cli

Next step would be the possibility to define custom prompts, toolsets and contexts for specific re-occuring tasks, and these appearing as tools to the main agent. Example for such a thing: create_new_page. The prompt could describe the steps one needs to create the page. Then the main agent could simply delegate this as a well-defined task, without cluttering its own context with the operational details.

Thanks for building this! The tool shows a lot of promise. Coming from Claude Code, the core functionality feels solid - just needs some permission refinements to match enterprise use cases. This is based upon quickly skimming the current code.

High ROI feature requests:

• Pattern-based permissions - Bash(git:) to allow git but not rm, Write(logs/.txt) for path scoping

• CLI permission flags - --allowedTools "Read,Bash(npm test)" --deniedTools "Write" for session overrides

• Allow/deny precedence rules - explicit deny should override general allow (security principle)

• Config file hierarchy - system → user → project precedence for policy enforcement

Medium ROI improvements:

• Command argument filtering - whitelist git commit but not git --exec-path=/bin/sh

• Multiple config formats - support both simple arrays and structured permission objects

• Runtime permission diagnostics - gemini permissions list to debug what's actually enabled

• Environment variable injection - top-level env config for OTEL endpoints, API keys, etc.

The permission engine is really the key piece - once you can express "allow X but not Y within X", it unlocks most advanced use cases. Keep up the great work!

One thing I'd really like to see in coding agents is this: As an architect, I want to formally define module boundaries in my software, in order to have AI agents adhere to and profit from my modular architecture.

Even with 1M context, for large projects, it makes sense to define boundaries These will typically be present in some form, but they are not available precisely to the coding agent. Imagine there was a simple YAML format where I could specify modules and where they can be found in the source tree, and the APIs of other modules it interacts with. Then it would be trivial to turn this into a context that would very often fit into 1M tokens. When an agent decides something needs to be done in the context of a specific module, it could then create a new context window containing exactly that module, effetively turning a large codebase into a small codebase, for which Gemini is extraordinarily effective.

So, as a member of an organization who pays for google workspace with gemini, I get the message `GOOGLE_CLOUD_PROJECT environment variable not found. Add that to your .env and try again, no reload needed!`

At the very least, we need better documentation on how to get that environment variable, as we are not on GCP and this is not immediately obvious how to do so. At the worst, it means that your users paying for gemini don't have access to this where your general google users do.

Google Gemini Google Gemini Ultra AI Studio Vertex AI Notebook LLM Jules

All different products doing the sameish thing. I don’t know where to send users to do anything. They are all licensed differently. Bonkers town.

Hi - I integrated Apple Container on M1 to run[1] the code generated by Gemini CLI. It works great!

1. CodeRunner - https://github.com/BandarLabs/coderunner/tree/main?tab=readm...

Hi, Thanks for this work.

currently it seems these are the CLI tools available. Is it possible to extend or actually disable some of these tools (for various reasons)?

> Available Gemini CLI tools:

    - ReadFolder
    - ReadFile
    - SearchText
    - FindFiles
    - Edit
    - WriteFile
    - WebFetch
    - ReadManyFiles
    - Shell
    - Save Memory
    - GoogleSearch

Thanks so much for this! I’d really appreciate a more consumer oriented subscription offering, similar to Claude Max, that combines Gemini CLI (with IP compliance) and the Gemini app (extra points for API access too!).

Requests for documentation clarification:

- Here [1] it says "Project settings override user settings." How does gemini determine if we're in a project? Does it look for a `.gemini` folder in the current working directory as well as every parent directory up to the root? Would Gemini be able to read the contents of a subfolder of the CWD if the subfolder contains a different `.gemini` folder?

- I don't see documentation for the `selectedAuthType` field in the documentation for settings.json. Mine says `oauth-personal`. I could've sworn I signed in with my Google Workspace account. Does `oauth-personal` apply to Workspace accounts?

And a feature request: it would be nice to add a restriction in the settings.json file forcing anybody who uses gemini in that project to sign in to a Workspace account in a specific org (or use a specific project, I guess).

[1]: https://github.com/google-gemini/gemini-cli/blob/main/docs/c...

Right now authentication doesn't work if you're working on a remote machine and try to authenticate with Google, FYI. You need an alternate auth flow that gives the user a link and lets them paste a key in (this is how Claude Code does it).

I have been evaluating other tools like Amp (from Sourcegraph) and when trying Gemini Cli on VsCode I found some things to improve:

- On a new chat I have to re-approve things like executing "go mod tidy", "git", write files... I need to create a new chat for each feature, (maybe an option to clear the current chat on VsCode would work)

- I have found some problems with adding some new endpoint on an example Go REST server I was trying it on, it just deleted existing endpoints on the file. Same with tests, it deleted existing tests when asking to add a test. For comparison I didn't find these problems when evaluating Amp (uses Claude 4)

Overall it works well and hope you continue with polishing it, good job!!

Using the Gemini CLI the first thing I tried to do was "Create GEMINI.md files to customize your interactions with Gemini." The command ran for about a minute before receiving a too many requests error.

> You exceeded your current quota, please check your plan and billing details. For more information on this error, head to: https://ai.google.dev/gemini-api/docs/rate-limits.

Discouraging

Feature request! :)

I'm a Gemini Pro subscriber and I would love to be able to use my web-based chat resource limits with, or in addition to, what is offered here. I have plenty of scripts that are essentially "Weave together a complex prompt I can send to Gemini Flash to instantly get the answer I'm looking for and xclip it to my clipboard", and this would finally let me close the last step in that scripts.

Love what I'm seeing so far!

The thing that had me close it and go back to Claude Code immediately was how often Gemini CLI failed to make file changes. This doesn’t instill confidence at all, with it retrying edits multiple times - who knows what it ends up editing, then.

CC has this issue too, but way less often, and second shot almost always works.

Pointed it at a project directory and asked it to find and fix an intentionally placed bug without referencing any filenames. It seemed to struggle finding any file or constructing a context about the project unless specifically asked. FWIW, Claude Code tries to build an 'understanding' of the codebase when given the same prompt. For example, it struggled when I asked to "fix the modal logic" but nothing was specifically called a modal.

Is the recommendation to specifically ask "analyze the codebase" here?

Could Gemini CLI be used for PR reviews? For example, would you expect that asking Gemini to compare two git commit hashes and analyze the code changes for potential bugs/conciseness/adhesion to the idiomatic project style, etc. to work well?

Edit: I tried it. The setup was a breeze. I fed the CLI two git commit IDs and some light prompting on what to look for. It gave a reasonable response. I'll try on a real PR shortly.

I'm just a hobbyist, but I keep getting the error "The code change produced by Gemini cannot be automatically applied. You can manually apply the change or ask Gemini to try again". I assume this is because the service is being slammed?

Edit: I should mention that I'm accessing this through Gemini Code Assist, so this may be something out of your wheelhouse.

Is LiteLLM integration on the roadmap? I get that part of the motivation to build this is to funnel usage to Google models, but it would be nice to be able to go through OpenRouter and/or use local models too. Open source is nice, but open model configuration would be nicer :)

Hello, thanks for the work for finally having an analog to Claude Code.

A natural question to ask is, if in the near future, can Google One "Google AI Pro" subscribers have higher limits than what is offered for free users?

i have a Google AI Pro subscription - what kind of credits/usage/allowance do i get towards gemini cli ?

Overall it looks and feels good. I gave it a problem, like, what was it... to update website pages to use single layout, it wasn't trivial but it wasn't that hard. It burned through 7M tokens and like 20-25mins didn't accomplished it. I stopped it because I didn't want to waste more resources.

I think with better prompting on my end, as I have good experience with Gemini, this will be awesome. You probably could tweak a lot on your end as well, don't let it get stuck in cycles.

Hi. It is unclear from the README whether the free limits apply also when there's an API key found in the environment - not explicitly set for this tool - and there is no login requirement.

Is there a reason all workspace accounts need a project ID? We pay for gemini pro for our workspace accounts but we don't use GCP or have a project ID otherwise.

Please, for the love of God, stop your models always answering with essays or littering code with tutorial style comments. Almost every task devolves into "now get rid of the comments". It seems impossible to prevent this.

And thinking is stupid. "Show me how to generate a random number in python"... 15s later you get an answer.

How often did you use gemini-cli to build on gemini-cli?

Two weeks back I made a simple wrapper CLI app for GEMINI https://geni.dev.

I strongly believe in AI cli apps for devs, congrats.

Is there a way to instantly, quickly prompt it in the terminal, without loading the full UI? Just to get a short response without filling the terminal page.

like to just get a short response - for simple things like "what's a nm and grep command to find this symbol in these 3 folders". I use gemini alot for this type of thing already

Or would that have to be a custom prompt I write?

Does it have LSP (language server) support? How should I think of this as different from Aider?

it would be cool to work with my google ai pro sub

Hi cperry - do you have an idea of when Gemini 2.5 live API will make GA? I saw a new GA label was created last week.

Congrats on your success. May you all be promoted and may your microkitchens be stocked.

Why does this CLI don't support google's A2A?

Thank you for supporting Windows natively.

one piece of feedback. Please do neovim on top of vim or have a way to customize the editor beyond your list.

F for the TPUs.

Hey the interface on YouTube loads super slowly for me. The change appeared a few months ago. I'm not talking about the video streams, but the ajax loading of the UI. Whether it's opening a new youtube tab or navigating between videos within youtube, it takes forever. Chrome/Safari -> same deal, 30 seconds delays is what I observe. My macbook pro is 10 years old, the problem doesn't appear on more recent hardware, but still youtube shouldn't be the slowest website to load on my machine. I can load spotify.com just fine in about 5 seconds.

Gemini has some fun failure modes. It gets "frustrated" when changes it makes doesn't work, and replies with oddly human phrases like "Well, that was unexpected" and then happily declares that (I see the issue!) "the final tests will pass" when it's going down a blind alley. It's extremely overconfident by default and much more exclamatory without changing the system prompt. Maybe in training it was taught/figured out that manifesting produces better results?

I asked it to do a comparatively pedestrian task: write a script to show top 5 google searches.

First it did the search itself and then added "echo" for each of them - cute

Then it tried to use pytrends which didn't go anywhere

Then it tried some other paid service which also didn't go anywhere

Then it tried some other stuff which also didn't go anywhere

Finally it gave up and declared failure.

It will probably be useful as it can do the modify/run loop itself with all the power of Gemini but so far, underwhelming.

This was also my exact experience. I was pretty excited because I usually use Gemini Pro 2.5 when Claude Code gets stuck by pasting the whole code and asking questions and it was able to get me out of a few pickles a couple of times.

Unfortunately the CLI version wasn't able to create coherent code or fix some issues I had in my Rust codebase as well.

Here's hope that it eventually becomes great.

Claude will do the same start over if things get too bad. At least I've seen it when its edits went haywire and trashed everything.

Same here. Tried to implement a new feature on one of our apps to test it. It completely screwed things up. Used undefined functions and stuff. After a couple of iterations of error reporting and fixing I gave up.

Claude did it fine but I was not happy with the code. What Gemini came up with was much better but it could not tie things together at the end.

Personally my theory is that Gemini benefits from being able to train on Googles massive internal code base and because Rust has been very low on uptake internally at Google, especially since they have some really nice C++ tooling, Gemini is comparatively bad at Rust.

So far I've found Gemini CLI is very good at explaining what existing code does.

I can't say much about writing new code though.

I tried it too, it was so bad. I got the same "revert" behaviour after only 15 minutes.

I wonder how much of this had to do with the context window size? Gemini’s window is 5x larger than Cladue’s.

I’ve been using Claude for a side project for the past few weeks and I find that we really get into a groove planning or debugging something and then by the time we are ready to implement, we’ve run out of context window space. Despite my best efforts to write good /compact instructions, when it’s ready to roll again some of the nuance is lost and the implementation suffers.

I’m looking forward to testing if that’s solved by the larger Gemini context window.

Current best practice for Claude Code is to have heavy lifting done by Gemini Pro 2.5 or o3/o3pro. There are ways to do this pretty seamlessly now because of MCP support (see Repo Prompt as an example.) Sometimes you can also just use Claude but it requires iterations of planning, integration while logging everything, then repeat.

I haven't looked at this Gemini CLI thing yet, but if its open source it seems like any model can be plugged in here?

I can see a pathway where LLMs are commodities. Every big tech company right now both wants their LLM to be the winner and the others to die, but they also really, really would prefer a commodity world to one where a competitor is the winner.

If the future use looks more like CLI agents, I'm not sure how some fancy UI wrapper is going to result in a winner take all. OpenAI is winning right now with user count by pure brand name with ChatGPT, but ChatGPT clearly is an inferior UI for real work.

Ask the AI to document each module in a 100-line markdown. These should be very high level, don't contain any detail, but just include pointers to relevant files for AI to find out by itself. With a doc as the starting point, AI will have context to work on any module.

If the module just can't be documented in this way in under 100 lines, it's a good time to refactor. Chances are if Claude's context window is not enough to work with a particular module, a human dev can't either. It's all about pointing your LLM precisely at the context that matters.

I feel like you get more mileage out of prompt engineering and being specific... not sure if "fix all the bugs" is an effective real-world use case.

Yeah but this collapses under any real complexity and there is likely an extreme amount of redundant code and would probably be twice as memory efficient if you just wrote it yourself.

Im actually interested to see if we see a rise in demand for DRAM that is greater than usual because more software is vibe coded than being not, or some form of vibe coding.

Yeah, and it's variable, can happen at 250k, 500k or later. When you interrogate it; usually the issue comes to it being laser focused or stuck on one specific issue, and it's very hard to turn it around. For the lack of the better comparison it feels like the AI is on a spectrum...

Claude seems to have trouble with extracting code snippets to add to the context as the session gets longer and longer. I've seen it get stuck in a loop simply trying to use sed/rg/etc to get just a few lines out of a file and eventually give up.

At some point, LLMs just get distracted and you might be better off throwing it away and restarting hah.

IIRC Claude Code indexes your code via a vector DB. That's why it's good for larger codebases. I'm not aware that GeminiCLI has this, too.

It's a lot more nuanced than that. If you use the free edition of Code Assist, your data can be used UNLESS you opt out, which is at the bottom of the support article you link to:

"If you don't want this data used to improve Google's machine learning models, you can opt out by following the steps in Set up Gemini Code Assist for individuals."

and then the link: https://developers.google.com/gemini-code-assist/docs/set-up...

If you pay for code assist, no data is used to improve. If you use a Gemini API key on a pay as you go account instead, it doesn't get used to improve. It's just if you're using a non-paid, consumer account and you didn't opt out.

That seems different than what you described.

To be honest this is by far the most frustrating part of the Gemini ecosystem, to me. I think 2.5 pro is probably the best model out there right now, and I'd love to use it for real work, but their privacy policies are so fucking confusing and disjointed that I just assume there is no privacy whatsoever. And that's with the expensive Pro Plus Ultra MegaMax Extreme Gold plan I'm on.

I hope this is something they're working on making clearer.

Hey all, This is a really great discussion, and you've raised some important points. We realize the privacy policies for the Gemini CLI were confusing depending on how you log in, and we appreciate you calling that out.

To clear everything up, we've put together a single doc that breaks down the Terms of Service and data policies for each account type, including an FAQ that covers the questions from this thread.

Here’s the link: https://github.com/google-gemini/gemini-cli/blob/main/docs/t...

Thanks again for pushing for clarity on this!

There is some information on this buried in configuration.md under "Usage Statistics". They claim:

*What we DON'T collect:*

- *Personally Identifiable Information (PII):* We do not collect any personal information, such as your name, email address, or API keys.

- *Prompt and Response Content:* We do not log the content of your prompts or the responses from the Gemini model.

- *File Content:* We do not log the content of any files that are read or written by the CLI.

https://github.com/google-gemini/gemini-cli/blob/0915bf7d677...

Mozilla and Google provide an alternative called gemmafile which gives you an airgapped version of Gemini (which Google calls Gemma) that runs locally in a single file without any dependencies. https://huggingface.co/jartine/gemma-2-27b-it-llamafile It's been deployed into production by 32% of organizations: https://www.wiz.io/reports/the-state-of-ai-in-the-cloud-2025

This is just for free use (individuals), for standard and enterprise they don't use the data.

Which pretty much means if you are using it for free, they are using your data.

I don't see what is alarming about this, everyone else has either the same policy or no free usage. Hell the surprising this is that they still let free users opt-out...

They really need to provide some clarity on the terms around data retention and training, for users who access Gemini CLI free via sign-in to a personal Google account. It's not clear whether the Gemini Code Assist terms are relevant, or indeed which of the three sets of terms they link at the bottom of the README.md apply here.

Kinda a tragedy of the commons situation. Everyone wants to use these tools that must be trained on more and more code to get better, but nobody wants it to be trained on their code. Bit silly imo.

Do you honestly believe that the opt-out by Anthropic and Cursor means your code won't be used for training their models? Seems likely that they would rather just risk taking a massive fine for potentially solving software development than to let some competitor try it instead.

Insane to me there isn’t even an asterisk in the blog post about this. The data collection is so over the top I don’t think users suspect it because it’s just absurd. For instance Gemini Pro chats are trained on too.

If this is legal, it shouldn’t be.

Is this significantly different than what we agree to when we put code on GitHub?

Some of my code is so bad I'm sure it will damage their models!

seems to be straight "yes" with no opt-out. https://github.com/google-gemini/gemini-cli/blob/main/docs/t...

How does that compare to Claude Code? How protected are you when using CC?

>If you use this, all of your code data will be sent to Google.

Not if you pay for it.

I'm not that bothered. Most of it came from Google or anthropic anyway

good to know, I won't be using this. Curious, do you know if OpenAI codex and Claude also do the same? I was under the impression that they don't share code.

Who cares, software has no value anymore.

Google suffers from Microsoft's issues: it has products for almost everything, but its confusing product messaging dilutes all the good things it does.

I like Gemini 2.5 Pro, too, and recently, I tried different AI products (including the Gemini Pro plan) because I wanted a good AI chat assistant for everyday use. But I also wanted to reduce my spending and have fewer subscriptions.

The Gemini Pro subscription is included with Google One, which is very convenient if you use Google Drive. But I already have an iCloud subscription tightly integrated with iOS, so switching to Drive and losing access to other iCloud functionality (like passwords) wasn’t in my plans.

Then there is the Gemini chat UI, which is light years behind the OpenAI ChatGPT client for macOS.

NotebookLM is good at summarizing documents, but the experience isn’t integrated with the Gemini chat, so it’s like constantly switching between Google products without a good integrated experience.

The result is that I end up paying a subscription to Raycast AI because the chat app is very well integrated with other Raycast functions, and I can try out models. I don’t get the latest model immediately, but it has an integrated experience with my workflow.

My point in this long description is that by being spread across many products, Google is losing on the UX side compared to OpenAI (for general tasks) or Anthropic (for coding). In just a few months, Google tried to catch up with v0 (Google Stitch), GH Copilot/Cursor (with that half-baked VSCode plugin), and now Claude Code. But all the attempts look like side-projects that will be killed soon.

Actually, that's the reason a lot of startups and solo developers prefer non-Google solutions, even though the quality of Gemini 2.5 Pro is insanely high. The Google Cloud Dashboard is a mess, and they haven't fixed it in years. They have Vertex that is supposed to host some of their models, but I don't understand what's the difference between that and their own cloud. And then you have two different APIs depending on the level of your project: This is literally the opposite of what we would expect from an AI provider where you start small and regardless of the scale of your project, you do not face obstacles. So essentially, Google has built an API solution that does not scale because as soon as your project gets bigger, you have to switch from the Google AI Studio API to the Vertex API. And I find it ridiculous because their OpenAI compatible API does not work all the time. And a lot of tools that rely on that actually don't work.

Google's AI offerings that should be simplified/consolidated:

- Jules vs Gemini CLI?

- Vertex API (requires a Google Cloud Account) vs Google AI Studio API

Also, since Vertex depends on Google Cloud, projects get more complicated because you have to modify these in your app [1]:

``` # Replace the `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION` values # with appropriate values for your project. export GOOGLE_CLOUD_PROJECT=GOOGLE_CLOUD_PROJECT export GOOGLE_CLOUD_LOCATION=global export GOOGLE_GENAI_USE_VERTEXAI=True ```

[1]: https://cloud.google.com/vertex-ai/generative-ai/docs/start/...

Google is fumbling the bag so badly with the pricing.

Gemini 2.5 Pro is the best model I've used (even better than o3 IMO) and yet there's no simple Claude/Cursor like subscription to just get full access.

Nevermind Enterprise users too, where OpenAI has it locked up.

Anthropic is the same. Unless it has changed within the last few months, you can subscribe to Claude but if you want to use Claude Code it'll come out of your "API usage" bucket which is billed separately than the subscription.

Some jerk has learned that we prefer CLI things and has come to the conclusion that we should therefore pay extra for them.

Workaround is to use their GUI with some MCPs but I dislike it because window navigation is just clunky compared to terminal multiplexer navigation.

There's also $300/mo AI ULTRA membership. It's interesting. Google One memberships even can't detail what "extra features" I can have, because it possibly changes every hour or so.

I actually had this exact same question when I read the docs, made an issue about it:

https://github.com/google-gemini/gemini-cli/issues/1427

That's valuable feedback and we're taking it to heart.

I was having such a good time using Gemini 2.5 Pro I forgot that Google is generally incapable of sustaining a new product. I’ll try not to get used too attached to it.

You don't get API keys for that subscription because it's a flat monthly cost.

And they say our scale up is siloed. Leave it to google to show' em.

You’d think with all this AI tooling they'd be able to organize better, but I think that the AI Age will be a very messy one with messaging and content

> Delightful!

You clearly have never had the "pleasure" to work with a Google product manager.

Especially the kind that were hired in the last 15-ish years.

This type of situation is absolutely typical, and probably one of the more benign thing among the general blight they typically inflict on Google's product offering.

The cartesian product of pricing options X models is an effing nightmare to navigate.

The other day I was trying to figure out a cost estimate for Gemini 2.5 Pro but they don't have this in their pricing calculator - it just isn't there. Only endless tables with various model sizes and options which aren't explained.

I think Pro is for regular folks, not specifically for programmers.

I also have a pro subscription and wish I could get an API key with that with generous quota as well but pro is just for "consumers" using Gemini app I guess

I was just trying to figure out if I get anything as a pro user. Thank you, you answered my question.

This is very confusing how they post about this on X, you would think you get additional usage. Messaging is very confusing.

Google needs to package their products better. It's always been the case with them. GCP is confusing because you can't easily find their services.

Google does not care about users at all, ZERO, it is full of people are so fond of technology, they try something and soon move on to something else. I am using Google Cloud for work, and previously Google Mail , Google Map, and YouTube user, zero good experience.

They are horrendous about their product strategy and positioning.

How can I subscribe to this? From the comments here it sounds like it's NOT bundled into the Gemini Pro subscription, but into "Gemini Code Assist Standard".

I can't find any way to upgrade to a paid plan, is this even possible for individuals, or is it just "free or enterprise"?

/Edit: Okay I went through the Gemini docs. I found that in Google Cloud you can enable Gemini Code Assist Standard and Enterprise for the account

- Standard is $19.00/mo

- Enterprise is $45.00/mo

Difference between the 2 editions: https://cloud.google.com/products/gemini/pricing

/Edit2: Found the actual management screen: https://codeassist.google.com/overview

I had a conversation with Copilot about Copilot offerings. Here's what they told me:

If I Could Talk to Satya...

I'd say:

“Hey Satya, love the Copilots—but maybe we need a Copilot for Copilots to help people figure out which one they need!”

Then I had them print out a table of Copilot plans:

- Microsoft Copilot Free - Github Copilot Free - Github Copilot Pro - Github Copilot Pro+ - Microsoft Copilot Pro (can only be purchased for personal accounts) - Microsoft 365 Copilot (can't be used with personal accounts and can only be purchased by an organization)

Because Gemini CLI is OSS, you can also find the system prompt at: https://github.com/google-gemini/gemini-cli/blob/4b5ca6bc777...

Funny how it suggests using the long-abandoned and recently deprecated create-react-app. Why not vibe code your app built on tech debt from day one!

It says to only use absolute paths but the temp file example uses relative. Nice.

You must be a busy man! Always new tools to review. How did you get interested in doing this?

Projects like this have to update frequently, having a mechanism like npm or pip or whatever to automatically handle that is probably easier. It's not like the program is doing heavy lifting anyway, unless you're committing outright programming felonies there shouldn't be any issues on modern hardware.

It's the only argument I can think of, something like Go would be goated for this use case in principle.

Ask Gemini CLI to re-write itself in your preferred language

This isn't about quality products, it's about being able to say you have a CLI tool because the other ai companies have one

Meanwhile, https://analyticsindiamag.com/global-tech/openai-is-ditching...

I really don't mind either way. My extremely limited experience with Node indicates they have installation, packaging and isolation polished very well.

Looks like you could make a standalone executable with Bun and/or Deno:

https://bun.sh/docs/bundler/executables

https://docs.deno.com/runtime/reference/cli/compile/

Note, I haven't checked that this actually works, although if it's straightforward Node code without any weird extensions it should work in Bun at least. I'd be curious to see how the exe size compares to Go and Rust!

> and doesn't require you to install a runtime like Node.

My exact same reaction when I read the install notes.

Even python would have been better.

Having to install that Javascript cancer on my laptop just to be able to try this, is a huge no.

Bun can compile to a single executable. Not sure if node has the same feature. Point is it's very doable with JS.

Language choice is orthogonal to distribution strategy. You can make single-file builds of JavaScript (or Python or anything) programs! It's just a matter of packaging, and there are packaging solutions for both Bun and Node. Don't blame the technology for people choosing not to use it.

You mean something like this:

https://www.npmjs.com/package/pkg

or perhaps this one:

https://www.npmjs.com/package/nexe

Node can also produce a single binary executable: https://nodejs.org/api/single-executable-applications.html

My thoughts exactly. Neither Rust not Go, not even C/C++ which I could accept if there were some native OS dependencies. Maybe this is a hint on who could be its main audience.

See gemmafile which gives you an airgapped version of gemini (which google calls gemma) that runs locally in a single file without any dependencies.

https://huggingface.co/jartine/gemma-2-27b-it-llamafile

Have you checked the https://github.com/google-gemini/gemini-cli/blob/main/docs/c... ? It has a section for workspace accounts.

It seems that you need to set up an env variable called GOOGLE_CLOUD_PROJECT https://github.com/google-gemini/gemini-cli/issues/1434

... and other stuff.

Same here.

When I was using it in cursor recently, I found it would break imports in large python files. Claude never did this. Do you have any weird issues using Gemini? I’m excited to try the cli today

what's your workflow?

Same thing. Couldn't get it to work with an API key (which worked just fine in Aider before). Couldn't get it to work with a Gemini Code Assist Standard license. The management UI and documentation are a torture maze. Can't stop thinking about all the better ways I could have used two our hours of my life.

This should be resolved: https://github.com/google-gemini/gemini-cli/discussions/2064

Got the exact same thing here, using Gemini with an API key.

Same here, I also have a env key which I use with Aider. Mind you aider+gemini has gone excruciatingly slow these days.

:) thank you for discovering our extensive documentation on this valued feature

I will never understand this kind of snark comment.

What are they supposed to do?

“Oh no, they’ve released CLI tool before us! It’s game over, we can’t do it too, we need to come up with something else now!”

I find it hard to imagine that any of the major model vendors are suffering from demand shortages right now (if that's what you mean?)

If you mean: This is "inspired" by the success of Claude Code. Sure, I guess, but it's also not like Claude Code brought anything entirely new to the table. There is a lot of copying from each other and continually improving upon that, and it's great for the users and model providers alike.

Yea, i'm not even really interested in Gemini atm because last i tried 2.5 Pro it was really difficult to shape behavior. It would be too wordy, or offer too many comments, etc - i couldn't seem to change some base behaviors, get it to focus on just one thing.

Which is surprising because at first i was ready to re-up my Google life. I've been very anti-Google for ages, but at first 2.5 Pro looked so good that i felt it was a huge winner. It just wasn't enjoyable to use because i was often at war with it.

Sonnet/Opus via Claude Code are definitely less intelligent than my early tests of 2.5 Pro, but they're reasonable, listen, stay on task and etc.

I'm sure i'll retry eventually though. Though the subscription complexity with Gemini sounds annoying.

And since they have essentially unlimited money they can offer a lot for free/cheaply, until all competitors die out, and then they can crank up the prices

The positive side is, there's a ton of free offerings you can pick up. Google workspace has one, ai studio has, deepmind etc etc

I think AI studio is the best form factor and cheapest?

Meanwhile,

Microsoft Copilot

Microsoft 365 Copilot

Copilot Pro

Microsoft Security Copilot

Microsoft Copilot Studio

GitHub Copilot

Copilot+ PC

Microsoft 365 Copilot for Sales

Microsoft 365 Copilot for Service

Microsoft 365 Copilot for Finance

This sounds like it's practically unlimited and free? A bit like GMail. It is comes to close to Claude Code they may overtake them and dominate the market simply with their tried and true "give it away for free forever with practially no limits" strategy. I pay 20$ for Claude Code, which is nothing in relation to the value it provides, but obviously a 0$ price point will get you a bigger piece of the Pie. Big Oof for Anthropic. Kinda lame that they can just copy a new and innovative company and undersell at a price point of 0$. Not that I'm cheering for Anthropic either; I'd much rather own my sillicon.

I agree that we should standardize on AGENTS.md but Gemini CLI does at least let you configure the filename: https://github.com/google-gemini/gemini-cli/blob/3a369ddec3b...

Claude Code seems to require the CLAUDE.md filename.

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

Hmm kinda makes sense to keep them separate because the agents perform differently, right?

You might want to tell Claude not to write so many comments but you might want to tell Gemini not to reach for Kotlin so much, or something.

A unified approach might be nice, but using the same prompt for all of the LLM "coding tools" is probably not going to be as nice as having prompts tailored for each specific tool.

Agreed. If there are provider-specific differences then we should be able to add CLAUDE/GEMINI.md to take precedence, but most things in there are basic or general information and should apply to all providers (same for .cursor/, .windsurfrules/, ..)

I think this whole thing sucks.. we should have a single way of setting these instructions that all agents will read. not sure why they feel they need to own that part. seems so petty. do you think something like this would be useful? https://github.com/luisrudge/dot-ai/ basically a cli that generates files for multiple AI tools based on a single source

They recently discontinued the main Gemini free tier which offered similar limits. I would say expect this to disappear when it hits GA or if it gets a lot of targeted abuse.

Because Google is rich and they'd like to get you hooked. Just like how ChatGPT has a free tier.

Also they can throttle the service whenever they feel it's too costly.

My guess: So that they can get more training data to improve their models which will eventually be subscription only.

Also I have Gemini Pro, what am I paying for?

I just tried it - it does not last long.

I think it was with the anthropic guys on Dwarkeshs podcast, but really it could have been any of the other tech podcasts with a big name AI guest, but anyway they were talking about how orgs need to make big decisions about compute allocation.

If you need to do research, pre-training, RLHF, inference for 5-10 different models across 20 different products, how do you optimally allocate your very finite compute? Weight towards research and training for better future models, or weigh towards output for happier consumers in the moment?

It would make sense that every project in deepmind is in constant war for TPU cycles.

> but it had only analysed a few files in the root and didn't generate anything at all

This is my experience with ALL AI editors and models. Half of the time, they say they changed things, but they didn't change anything.

its share data ON by default and you can turn it off in the options

In my experience the longer thinking time before acting is due to the model (2.5 Pro). Claude Sonnet 4 is just snappier overall. Same experience in tools like Cursor when comparing these models.

Similar. Yesterday things seemed to be going okay. It was trucking along, making code changes.

Then I hit the rate limit. - Fine, no worries, it'll be interesting to see if the quality changes.

Then it starts getting stuck and taking forever to complete anything. So I shut it down for the day.

Today, I start it back up and ask it to pickup where it left off and it starts spinning. I forget about it and come back 7.5 hours later and it' still spinning. When I kill it it said: 1 Turn, 90k input tokens, 6.5 hours of API time... WTH?

And now I'm just totally rate limited - `status: 429, statusText: 'Too Many Requests'` - every time. Also, I can't find any kind of usage data anywhere!

I got the same experience. I ran `gemini --debug` and it spit out the authentication URL at least. However I got the same `Rate limiting detected` for a simple refactoring task in a repo with 2 python files after a few minutes.

not working for me either. getting "too many requests". my own CLI tool i wrote for gemini last week works better lol. maybe i have something configured wrong? though that really shouldn't be the case. my api key env var is correct.

I have nothing against npm, but having a single binary would've been a lot more trivial and convenient. Codex is heading in that direction, and I hope others do too.

Gemini web UI is extremely apologetic and self deprecating at times.

Therefore I was not surprised to experience Gemini spiraling into an infinite loop of self-deprecation - literally it abandoned the first command and spiraled into 5-10line blocks of "i suck"

---

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.

Maybe this will help: https://github.com/armanirodriguez/ansifade-py

And it is a blocky gradient, each character is a color. It's just the gradient they chose is slow enough that you don't notice.

Look at the package.json inside the CLI folder https://www.npmjs.com/package/ink-gradient

we went through several versions of this, it was the second most fun part after authoring the witty loading phrases.

It's firmly on the radar - we will have a great answer for this soon.

Or aider. In any case, while top llms will likely remain proprietary for some time, there is no reason for these tools to be closed source or tied to a particular llm vendor.

I've had a good experience with https://kilocode.ai

It integrates with VS Code, which suits my workflow better. And buying credits through them (at cost) means I can use any model I want without juggling top-ups across several different billing profiles.

Can it use Claude sub or Gemini free tier the same way Gemini CLI does?

How does its CLI mode compare to Claude Code and Gemini CLI?

opencode.ai

What’s neat is we can proxy requests from Gemini or fork it with only replacing the api call layer so it can be used with local models!!!

I love healthy competition that leads to better use experiences

As a heavy Claude code user that's not really a selling point for me

Ultimately quality wins out with LLMs. Having switched a lot between openai, google and Claude, I feel there's essentially 0 switching cost and you very quickly get to feel which is the best. So until Claude has a solid competitor I'll use it, open source or not

No per-token billing here either: "...we offer the industry’s largest allowance: 60 model requests per minute and 1,000 requests per day at no charge."

https://blog.google/technology/developers/introducing-gemini...

If you use your personal gmail account without billing enabled, you get generous requests and never have to worry about a surprise bill.