Entropic just added a bunch of features

to clot code and clot desktop including

the ability of clot to ask you

proactively for clarifications.

Now you can use clot code directly on

the web very similar to codec cla and

their web interface. They also released

some pretty neat updates to their

desktop application with some pretty

awesome features which are going to make

it extremely useful. They also released

a new sandbox and actually made it open

source which is very unusual for

Enthropic. So we're going to look at all

these updates and a lot more in this

video. Let's first talk about the latest

updates to cloud desktop which is now

generally available. So when you install

cloud desktop it's going to be available

in your menu bar. You can also access it

through a shortcut which is tapping

options twice on MacOSS. So that brings

up a quick view of Cloud Desktop. But

here's the neat feature. You can

actually just drag and select an area

that is going to be sent to Cloud

Desktop and then we can tell it to do

something. Can you analyze what is in

these images and explain those to me?

Which is actually pretty neat,

especially if you want to use claude as

a helpful assistant. Now, it can see

your desktop. But there is also another

pretty neat feature which also has a

very weird implementation. So, let me

show you that. This next one is talk to

clone. So you can trigger talk to clone

just by pressing the caps lock button

which was a very weird choice. Here is

how it looks like. Now if you start

talking Claude is going to transcribe

the text and then if you press on caps

lock again it will stop transcribing.

Okay. So this is pretty neat. Although I

think it did not capture the whole

thing. But here's the weird part. Now,

if I'm here and I press caps locks

again, it doesn't do anything, right?

So, it's a very weird experience. Now, I

have a much better option for you. You

can actually use my application that I'm

calling, right? So, this gives you the

ability to transcribe your speech into

any text box. There are a number of

different short keys. It's not caps

lock. So, let's bring up cloud desktop

again. And now we're going to press the

short keys. I capture this image and now

I can trigger this app, perform image

analysis and explain what exactly you

see here. Now the beauty is it's

extremely fast and it works in any text

window inside the Mac OS. Details are

going to be in the video description if

you are interested in this app. It's a

perfect companion for wbe coding or

writing purposes and it's going to save

you so much time. I use it daily. This

next feature is something that a lot of

people have been asking for and that's

the ability to use cloud code on the

web. Very similar to codeex where you

have a version that you can use on the

web and you also have the codec cla.

Now, cloud code also gives you the

ability to continue working on this web

interface where you can assign the agent

a task and then it's going to create a

PR for you. So, let me show you how this

works. So, here is the claw code version

of this asynchronous agent. Now, you'll

have to connect your GitHub account. You

can see that it already completed a task

for me. So, it's ready to create a PR.

But let me walk you through what the

flow actually looks like. Okay. So we're

going to provide a very detailed prompt

on what exactly I want this to look

like. Now in here you can use

environment variable. So if you want to

set any API keys you can set them here.

And you can also define the network

access. So whether you want to use it on

trusted network access, no network

access at all or custom access. Right?

In this case, I don't really need any

environment variables or any API keys. I

already created new repo for this. We're

going to just send that in and cloud

code is going to start working. So, in

this case, I just wanted to create a

website for one of the apps. Now, you

can run multiple asynchronous agents.

So, you can potentially create yet

another task for cloud code and it's

going to continue working on it. Okay.

For some reason, I'm having a lot of

connection issues right now. I had to

actually archive one of the previous

attempt, but it did work in one of my

previous request. So, this might be

something related to the network right

now. Now, moving on. Now, cloud code is

pretty awesome, but it can be pretty

expensive to run it. So, people have

been trying openweight models and one of

my favorite is Kim K2. not only its

state-of-the-art model, but it actually

has a personality which is very pleasing

to talk to. So, I was really excited

when they reached out to sponsor this

video. It's a free alternative to

something like Chart TPD, but they also

recently released their agent mode which

they're calling okay computer. So, think

of this deep research meets image

generation meets lovable. From a single

prompt, you can create working web pages

and web apps. So, let me show you a

quick demo of this in action. So, here

we select okay computer. I'm going to

provide a very detailed prompt and let's

send this in. It sets up a virtual

machine and then start thinking about

the specs. It creates tasks for itself

and then goes working. So here's the

website that it created and there is a

real working link to it. Seems like

pretty neat and actually usable. So do

check it out. You get to try it for

free. Link is going to be in the video

description. Now back to the video. If

you're using cloud code in the normal

mode, it will keep asking you for your

permissions before it performs certain

operations, which is usually pretty

great, but can be extremely annoying.

Now, on the other hand, if you're

feeling adventurous, you can actually

use the yolo mode or skip permissions

dangerously. But this can perform some

weird operations on your system. That's

why claude just introduced something

called sandbox which actually sandbox is

the operations that it's going to

perform and you can limit where it can

do those operations. So for example if

you go to / sandbox right now we are in

no sandbox that means any operation

performed by claude can be outside the

scope of the current folder and it can

do system level operations because we

are in the yolo mode. But with sandbox

you can limit all the operation that are

going to be performed to a current

folder or set of folders and files.

Right? There are two different options.

One is sandbox patch tool with auto

allow in accept edit modes and then the

other one is sandbox with regular

permissions. So if you click on this is

going to basically sandbox all of the

operations to this specific folder but

it's going to still ask you for

permissions when it's doing an

operation. Now this means that every

cloud code instance is going to ship

with a sandbox that is going to be

running on your machine. So here's the

overall architecture. You have the user

interactions cloud code and then that

sandbox environment in which is

performing all the operations. In fact,

the cloud code web version is a sandbox

environment that tropic is running on

their own servers and in that case your

GitHub credentials are actually outside

of the sandbox environment. Now the good

news is that they opensource the

anthropic sandbox runtime environment.

So if you are building agents that needs

a sandbox, you can build on top of this.

I highly recommend to check this out.

And this actually solves a real problem

in terms of managing the access of the

agent to the system. And according to

their own research, they found that

sandboxing safely reduces permission

prompts by 84%. They can be annoying at

times. So I highly recommend to check

out this feature as well. Claude can now

proactively ask you questions if it

needs certain clarifications. So here's

a quick example. Now in this case I

specifically told it to ask me

clarifying questions. It can do it by

itself now. So for example I want to

implement a new feature in my speech

transcription app. I can select the

answers to the different questions right

and that's the issue right and then you

can submit. It doesn't happen often but

if it needs clarification it will

proactively asks you. Something that I

used to do was whenever I provide a

prompt, I would actually ask the LLM to

ask for clarification if it needs any.

But now with this, it's pretty awesome

and the questions that it asks are

extremely relevant and well thought out.

But as a user, you don't really have

control on it. It will proactively asks

you. The last thing we are going to look

at is cloud skills that you can now

directly use within cloud code. I

previously created a dedicated video on

what exactly claude skills are and

there's also a video coming out on how

to use them directly within cloud code.

Link to that is going to be in the video

description. If you go to capabilities

within the cloud desktop or the web

version, you can see there is a skill

called skill creator. You can use this

to create your own skills. So here's an

example. I uploaded all the ADK

documentations and just asked claude to

create a skill for that. So then it

started using the actual creator skill

and created a zip file. All you need to

do is to just download that zip file.

Then go to the project folder where you

want to use that skill within the cloud

folder. Create another skills folder and

just copy the skill folder there. And if

you want to use it, here's what you're

going to do. Can you use the ADK skill

to create a multi- aent system using the

Google agent development kit? This is

supposed to be a customer support agent.

Make sure that it is able to process

orders, returns, and give the user

information. Now if we send this in then

cloud code will be able to just go

through that skill and use the ADK

documentation to help us create an

agent. So here it's actually asking us

to use the ADK skill. So make sure to

subscribe to the channel to watch the

other video that is going to be coming

out on creating skills and using them

within cloud code. Now anyways I hope

you found this video useful. Thanks for

watching and as always, see you in the

next