Claude Cowork Expands to iPhone and the Web

[MacRumors]

Anthropic today said it is bringing Claude Cowork to mobile and the web, with beta access rolling out to Max users first. Claude Cowork is a feature that lets Claude access local files and connected tools, then autonomously complete tasks on your behalf.

Claude Cowork tasks can now be started and monitored on the web, desktop, and Claude mobile apps. Claude is able to work in the background in the cloud even when no device is online, sending an alert when it needs permission to move forward with a task.

Anthropic says the desktop is still the best place for the full Cowork experience because it gives Claude access to local files and a browser, but those who were unable to install a desktop app can now use Cowork.

Claude chat and Cowork have also been combined into a single view on the web and desktop and are no longer separate tabs.

As part of the expansion to mobile, Anthropic is extending doubled Cowork usage limits through August 5.

Article Link: Claude Cowork Expands to iPhone and the Web

 

[TwoBytes]

Even as a paid subscriber, I hit my Limit so much that Claude is a real black hole of money

 

[Slix]

"2 yours, 1 theirs" is such a bizarre phrasing of that type of information.

 

[uacd]

I mean, AI fatigue is a real thing…

 

[FoxyKaye]

“Access to local files and a browser” - No thank you.

 

[Astuces iOS]

Anthropic: We feel like we should open new ways to pollute, that's why we introduce cowork on the web and on the phone ! We want more people to pollute ! LET'S GO POLLUTION !

 

[contacos]

Lol I noticed it the other day and thought this means you can cowork on one AI agent chat with your colleagues and that's why I never even clicked on it.

 

[RumorConsumer]

****ing great!

 

[Mrkevinfinnerty]

Now you can wait for your rate limit to reset on mobile 🔥

 

[chuckles235]

And it doesn’t require thousands of dollars of hardware upgrades to work!

 

[Black_Mage]

Even as a paid subscriber, I hit my Limit so much that Claude is a real black hole of money

After reading this post, it just occurred to me that A.I. has the dangerous potential to become a very expensive addiction. Worse than gambling.

 

[jeroenvip]

Even as a paid subscriber, I hit my Limit so much that Claude is a real black hole of money

I'm a developer using the Claude 5X plan, five days a week for 8–10 hours a day. I'm currently building a rental platform that eliminates the 15% commission charged by most major booking platforms.

Over the past four months, I've hit the weekly usage limit about 40–50% of the time. Ironically, I'm also paying for features I never use, the 50% left over each week.

I've found that the key to success is writing a detailed, well-structured Markdown (MD) file—almost like a book. Every architectural decision, every feature, every workflow, and every implementation step is documented. Before the AI writes a single line of code, it already knows exactly what it's supposed to build.

You don't let the AI be the architect. You are the architect. The AI is simply an incredibly fast typist and implementation assistant.

Or are you the kind of developer who prompts something like:

"Build me a complete production-ready platform for Product X. Fix any bugs, don't make mistakes, and here's my entire codebase, including all modules and the vendor folder."

Those are two completely different ways of working with AI, and in my experience, the first approach consistently produces better results.

 

[UltimaKilo]

Anthropic really wants to make OpenClaw irrelevant with Claude. The final nail would be an on-device model that decides when it needs the cloud to bring down token usage. I bet this comes sometime next year.

 

[drumcat]

I still kind of feel this AI stuff should be a separate website. It's not really Apple related other than the app being available on the App Store as many others are as well.

It's very related. Something I wanted to do on my iPhone yesterday I can do today.

 

[rydewnd2]

Anthropic really wants to make OpenClaw irrelevant with Claude. The final nail would be an on-device model that decides when it needs the cloud to bring down token usage. I bet this comes sometime next year.

Tangential to your point: I think we’ll look back on the way we use AI today and laugh. Manual model selection is a terrible user experience and I doubt even the most advanced users of AI can legitimately and accurately predict when they should be using one model over another.

 

[drumcat]

Anthropic really wants to make OpenClaw irrelevant with Claude. The final nail would be an on-device model that decides when it needs the cloud to bring down token usage. I bet this comes sometime next year.

Anything to offload onto a local model would be welcomed. Imagine how much of an advantage it would be if you could have the cloud AI making decisions, but your willingness to slow down and be local saves you money. I really see this as the gold standard...

You have something like FluidVoice on device, GoldenGate acting as a traffic cop and MCP regulator, ClaudeLocal having what amounts to a node on your machine, and an integrated cloud textUI/conversationUI orchestrating. That's the biggest paradigm change since applications assuming network/internet-first.

 

[BarredOwl]

Bring it on. Great news.

 

[SBlue1]

I'm a developer using the Claude 5X plan, five days a week for 8–10 hours a day. I'm currently building a rental platform that eliminates the 15% commission charged by most major booking platforms.

Over the past four months, I've hit the weekly usage limit about 40–50% of the time. Ironically, I'm also paying for features I never use, the 50% left over each week.

I've found that the key to success is writing a detailed, well-structured Markdown (MD) file—almost like a book. Every architectural decision, every feature, every workflow, and every implementation step is documented. Before the AI writes a single line of code, it already knows exactly what it's supposed to build.

You don't let the AI be the architect. You are the architect. The AI is simply an incredibly fast typist and implementation assistant.

Or are you the kind of developer who prompts something like:


Those are two completely different ways of working with AI, and in my experience, the first approach consistently produces better results.

AI can be such a boost for creativity if you use it right. Instead of prompting it to help me write a book I asked ChatGPT to make me a five day training how to start writing. Yeah I could have searched for online courses but I wanted something customized to my needs, skipping things I know and focusing on things I don't know, and optimized for my speed of learning my existing experience and my available time.

So you are completely correct, you are the boss, you make the decisions and set the course, AI is just there to implement and execute the work for you. A lot of people use AI wrong and get bad results.

 

[UltimaKilo]

Tangential to your point: I think we’ll look back on the way we use AI today and laugh. Manual model selection is a terrible user experience and I doubt even the most advanced users of AI can legitimately and accurately predict when they should be using one model over another.

You know, I bet that happens within the next 18-24 months. It is so incredibly early that most people still don’t use LLMs, at most, they use them like an advanced search engine.

The next couple of years are really going to be insane as all these feature become better and more user friendly.

 

[UltimaKilo]

Anything to offload onto a local model would be welcomed. Imagine how much of an advantage it would be if you could have the cloud AI making decisions, but your willingness to slow down and be local saves you money. I really see this as the gold standard...

You have something like FluidVoice on device, GoldenGate acting as a traffic cop and MCP regulator, ClaudeLocal having what amounts to a node on your machine, and an integrated cloud textUI/conversationUI orchestrating. That's the biggest paradigm change since applications assuming network/internet-first.

Most people are unaware how efficient these new local models are going to be by late next year, whereby an optimized 9B model will routinely outperform today’s 32B models on the same hardware, delivering higher quality output with way less power draw, faster inference, and the ability to run smoothly on less memory thanks to continued gains in training data, architectures, and reasoning techniques.

That’s why I see so many people running local models panicked that they don’t have 48GB of RAM where a mere 24 GBs will likely serve them well for the next few years.

With Apple knowing how popular the Mac Mini M4 has been for people running local models, I hope they skip an M5 variant and going for M6 for the much higher memory bandwidth and the much better GPU and upgraded Neural Engine. Doesn’t seem like that is happening, though.

 

[drumcat]

Most people are unaware how efficient these new local models are going to be by late next year, whereby an optimized 9B model will routinely outperform today’s 32B models on the same hardware, delivering higher quality output with way less power draw, faster inference, and the ability to run smoothly on less memory thanks to continued gains in training data, architectures, and reasoning techniques.

That’s why I see so many people running local models panicked that they don’t have 48GB of RAM where a mere 24 GBs will likely serve them well for the next few years.

With Apple knowing how popular the Mac Mini M4 has been for people running local models, I hope they skip an M5 variant and going for M6 for the much higher memory bandwidth and the much better GPU and upgraded Neural Engine. Doesn’t seem like that is happening, though.

Same. I've said as much. The Mini M6, optimised for exactly this, is going to be Apple's best selling computer ever. Some will understand, but most will have to catch up.