It looks like the key idea here is that a GitHub repo can have a gemini-extension.json file which specifies some MCP servers and a GEMINI.md file, then you can tell Gemini CLI to add that repo and it will fetch and configure those details.
"The extensions listed here are sourced from public repositories and created by third-party developers. Google does not vet, endorse, or guarantee the functionality or security of these extensions. Please carefully inspect any extension and its source code before installing to understand the permissions it requires and the actions it may perform. "
Extensions are always a trust challenge, but the high value of AI systems means that I expect we'll see a very high volume of attacks in the near future.
My recent experience of Gemini CLI in comparison to Claude code is very negative. For a complex task of trying to migrate some code base from CPU-only to GPU computations Claude was really helpful, and allowed me to conduct a range of experiments and converge towards sensible architecture. In the same time I simply was not able to get anything useful out of Gemini. Some code was written, but it was not tested, was failing or timing out and after repeated prodding Gemini got into loops or internal errors.
Watching Gemini CLI repo daily is great insight into how Google works. Tons of commits, three release channels, great ambitions and very little to show. These guys are running in all directions and getting nowhere.
I once spent an hour trying to get an API key for Gemini from AIStudio and Vertex, with assistance from LLMs, and finally gave up and used OpenRouter. It's that great an experience.
I’ve regularly used prompts along the lines ‘please modify the UI tor x feature to look like the selected Figma frame’. It’s ideal for a designing developer who wants to be much more specific about the user experience.
If there's value in a human using Figma, there's value in being able to give a "bot" instructions to do whatever a person might. e.g. "Gemini, update all the Figma layouts with this new text."
Figma MCP doesn't seem very useful at all and seem to me to compete with their newer prompt based builder. I think they're probably have their AI fork in the road. If they can't integrate these tools someone else will...
Presumably the reverse would be more valuable: design something in Figma, then have Gemini CLI read from Figma and use that as a reference while writing the code.
I don't know if this is how the Figma integration works, but right now I'm just manually screenshotting my designs and passing them into Claude CLI as references for what I want, so this seems potentially more streamlined.
Cool, but at the same time, it feels overwhelming: so many different CLI or IDE tools, so many extension points. It will be fascinating to see how this all shakes out.
It's more the agentic stuff, and things specific to the gemini-cli, which is behind the alternatives in features and capabilities. It was also making an insane number of requests (>1000 in a day, which is about the same as my Copilot usage for a month), but I'm sure they are doing their accounting differently. Google has tried to do AI accounting differently, but has acquiesced to counting tokens instead of chars, fingers crossed they do here to instead of being a snowflake that takes more effort to align in comparisons
I can't stand the cutesiness they've embedded into it either. I don't want that in a work tool
My general sense is that ai-clis will lose out to IDE integrations. I'd prefer a single tool and experience over having to context switch. Putting my AI partner in the same tool and env I use is better than having it separate with hacks to make it seem like it can be in there too, only sorta not quite
I'd add my general impression is that all the good designers that were at Google are gone, looking across their portfolio
Take even the Gemini Web App... what set Google apart in the early days? Search was just an input box, no clutter or calls to action. They have recently decided to break from this (they did have it clean beforehand) and try to get me to use image generation and other calls to action. Please get rid of the slop before I can even make my own slop!
and don't egg me on about Google Cloud... it's speed now feels like Jira, which to many people's surprise has changed course and is quite fast now
Using both. Sometimes Claude makes stupid mistakes, such as skipping the provided document. Overall, Claude gives better enterprise-grade suggestions and code, which Gemini sometimes misses. However, if you have in-depth documentation of the scope and implementation approach, Gemini performs better.
My primary tool is Gemini Code Assist. Claude is used to create the draft implementation approach and for a final sanity check of the code, as well as to propose enhancements for production readiness. This combination has worked well for me. Since Claude is expensive and Gemini is more affordable, it also makes economical sense.
I usually provide a well defined scope and detailed implementation approach, with the whole project split into submodules and the scope and implementation approach is again refined for each modules. In my experience, the programming language also matters, results are often better when using statically typed languages. We use C# and front end is always developed without AI.
Gemini is about 10x cheaper per token. But for some reason it's using 8 times more input tokens than CC. They also have this thing called cached tokens, which is much cheaper than not cached tokens. It's a hot cache of your context on Google side, cached automatically.
So at the end of the day you don't know how much you'll pay.
Models
Google is good for very complex topics and when the conversation is short. But both models are great. I prefer Claude and Sonnet 4.5 is great all around
CLI tools
Gemini cli is at it's very early days. Doesn't support hooks or subagents. Often runs into loops it can't break out from, essentially gets stuck but you still pay for the tokens.
Claude is just great. Allows you to write complex workflows they way they are supposed to be written. Handles hooks and subagents. MD file can reference another MD file, so you can DRY your files.
Nested plan mode works weird, sometimes the agent gets stuck if it asks for plan approval and thinks it's executing it, but displays nothing... So plan mode is not fully supported in subagents.
A nice thing is that .Claude directory is automatically understood by codex or cursor, you should be able to run your Claude command using openai models via codex or maybe even other providers via Cursor.
Summary
Overall Claude is the best all around, but the tokens are crazy expensive and the subscription model is a joke. You don't know how many tokens you can use when you're subscribed, but it's 'something', and last week they changed the limits, it's suddenly half of 'something'...
Same with me, I had signed up for a year plan on Cursor, used it for like... a week, then tried CC and it was so much better I signed up for the max plan and then never looked back.
The last time I was locked out of Claude Code I gave gemini cli a try. 10 minutes and lots of weird scrolling and screen flashing later, and it kept trying to add and remove the same function from the same file over and over, in a loop, until it exhausted its credits and demanded I upgrade.
Here's a commit I found adding one of these for the Google Maps platform: https://github.com/googlemaps/platform-ai/commit/95c4efdb43c...
Extensions are always a trust challenge, but the high value of AI systems means that I expect we'll see a very high volume of attacks in the near future.
I used to use Gemini occasionally with cursor 6 months ago and it wasn’t this bad. So I’m not sure what exactly caused it to be that bad.
Why would I want to do that?
I don't know if this is how the Figma integration works, but right now I'm just manually screenshotting my designs and passing them into Claude CLI as references for what I want, so this seems potentially more streamlined.
This one seems to showcase a bunch of the "extension" features, including a custom MCP for dealing with file line numbers.
EDIT:
I've posted about it on GitHub: https://github.com/gemini-cli-extensions/security/issues/81
Hopefully the relevant team will see it there.
It's more the agentic stuff, and things specific to the gemini-cli, which is behind the alternatives in features and capabilities. It was also making an insane number of requests (>1000 in a day, which is about the same as my Copilot usage for a month), but I'm sure they are doing their accounting differently. Google has tried to do AI accounting differently, but has acquiesced to counting tokens instead of chars, fingers crossed they do here to instead of being a snowflake that takes more effort to align in comparisons
I can't stand the cutesiness they've embedded into it either. I don't want that in a work tool
My general sense is that ai-clis will lose out to IDE integrations. I'd prefer a single tool and experience over having to context switch. Putting my AI partner in the same tool and env I use is better than having it separate with hacks to make it seem like it can be in there too, only sorta not quite
Take even the Gemini Web App... what set Google apart in the early days? Search was just an input box, no clutter or calls to action. They have recently decided to break from this (they did have it clean beforehand) and try to get me to use image generation and other calls to action. Please get rid of the slop before I can even make my own slop!
and don't egg me on about Google Cloud... it's speed now feels like Jira, which to many people's surprise has changed course and is quite fast now
But nice of them to try wrestling a file into your repo named “gemini something something”.. can’t knock them for having a go.
I feel no desire to switch or learn a new thing but I'm wondering if people feel like it's on par with CC or Codex or behind.
My primary tool is Gemini Code Assist. Claude is used to create the draft implementation approach and for a final sanity check of the code, as well as to propose enhancements for production readiness. This combination has worked well for me. Since Claude is expensive and Gemini is more affordable, it also makes economical sense.
I usually provide a well defined scope and detailed implementation approach, with the whole project split into submodules and the scope and implementation approach is again refined for each modules. In my experience, the programming language also matters, results are often better when using statically typed languages. We use C# and front end is always developed without AI.
I use Gemini Standard Tier and Claude Pro Tier.
Gemini is about 10x cheaper per token. But for some reason it's using 8 times more input tokens than CC. They also have this thing called cached tokens, which is much cheaper than not cached tokens. It's a hot cache of your context on Google side, cached automatically. So at the end of the day you don't know how much you'll pay.
Models
Google is good for very complex topics and when the conversation is short. But both models are great. I prefer Claude and Sonnet 4.5 is great all around
CLI tools
Gemini cli is at it's very early days. Doesn't support hooks or subagents. Often runs into loops it can't break out from, essentially gets stuck but you still pay for the tokens.
Claude is just great. Allows you to write complex workflows they way they are supposed to be written. Handles hooks and subagents. MD file can reference another MD file, so you can DRY your files.
Nested plan mode works weird, sometimes the agent gets stuck if it asks for plan approval and thinks it's executing it, but displays nothing... So plan mode is not fully supported in subagents.
A nice thing is that .Claude directory is automatically understood by codex or cursor, you should be able to run your Claude command using openai models via codex or maybe even other providers via Cursor.
Summary
Overall Claude is the best all around, but the tokens are crazy expensive and the subscription model is a joke. You don't know how many tokens you can use when you're subscribed, but it's 'something', and last week they changed the limits, it's suddenly half of 'something'...
And codex seems to be catching up. But gemini-cli today was really bad on an elixir edit. I’m not sure what causes it to be that bad.
https://github.com/google-gemini/gemini-cli/issues/4310
Well I guess not...