Anthropic Just Added These Features to Claude Code

Okay, so a week later, we're back with even more Cloud Code updates, and I'll be going through many of the interesting ones that you can see in the list here. But before getting started, this video is sponsored by myself and my Cloud Code Masterass. It is the most comprehensive class on Cloud Code that you will find online. And nowadays, many companies are paying a premium for being able to use Cloud Code effectively and build natively with it. The new year sale is almost over and it will get more expensive over time, so it's better to

lock in the price now. Anyways, so the first thing that caught my eye is this brand new setting to decide where your planning directory is. So currently the default in cloud code is it basically stores your plans into the dotclaw directory in your computer in this like plans folder with these random names and it's not really useful to be able to then like inspect the plan or go back to it or even get another coding agent to look for the plan. But what you can now do in cloud code is that if you define the plans directory in your config folder then you can have it saved within the same directory. Then you can do other things like getting Gemini CLI or Codex CLI to inspect the plan or DI implementation or review or critique the plan because you can tag the file quite easily now. And if you store all your plans within your projects folder, then you can easily use a GP to then search for your plan history and then improve the performance of cloud code because it has some kind of like reference or knowledge of what it did previously. So to enable this, you want to go to yourcloud folder settings. json JSON and then type in something like this, the path where you want your plans to be stored and then making sure you're on the latest version by running cloud update and then running cloud. Can you give me a short example plan that basically says hello world for a YouTube video. And then you can see in the prompt over here to recognize that it should save the plan into plans folder. And then afterwards when I do yes and accept edits, then you can see that if I go to folder YouTube plans, I can see the exact same plan here. And now that we have the plan, we can tag it in a different agent like Codex CLI for example and basically get it to critique and improve on the plan. And you can also turn this into workflow using the exit plan mode post tool use hook. So Claude would write the plan, you would approve the plan and then the hook gets triggered automatically. And you could have one hook that would then save the plan into a dated folder with a better name for example and then you can have another hook where it would pass it automatically to codeci which is pretty good when it comes to planning or also Gemini CLI or some other CLI tools as well. Or you could do something like passing the plan into other instances of cloud code where it's running different sub aents that do different reviews. So you could have cloud code running as a security check sub agent, another running as like a cost analysis sub agent and that basically gives cloud code feedback on the plan for it to go through one more iteration and improve on the plan automatically. So to basically show you what this workflow would look like in reality, you can see I have the post tool use exit plan mode matcher and then I have two hooks being triggered. One of them is archive plan. So essentially what the archive plan does is it copies it into a dated folder which looks kind of like this with a timestamp and a session ID. So cloud code can refer back to previous plans and sessions quite easily. So you could do something like a genensic gripping across previous plans and sessions like I've covered in my cloud code masterass. And then I have a nova hook here that basically passes that plan into a headless version of codeci and gemini cli to get feedback on the plan before going through one more iteration and improving on the plan. So to basically show you what this looks like, can you add 11 lab streaming to the Mac OS application so they can use their API key to automatically stream words onto their computer? So pressing that cloud code will do the usual thing where it comes up with a plan. And something else that cla code did add to planning mode is when it asks you a question, you can either say chat about this or skip interview and plan immediately. Okay, now this plan is done and we have even more new features where it says would you like to proceed? You can do yes, clear context and auto accept edits. Yes, clay context, manual accept edit and then the two previous options which is yes o ex accept and manual accept in the same context window. So essentially what is happening here in ch number one when cla code was planning it spun up a bunch of explore sub agents to look for the codebase and establish patterns that can be used to make a better plan. Those exploration results came back to the main session. It may have done some web fetching or using MCP servers to find more things relevant. It did the planning with a planning sub aent and this is a result of the planning sub agent and then it has a final plan and you can see now we've already used up quite a bit of the context window. If I remember correctly this used up about 60 70,000 tokens so far and if the plan ends up taking quite a lot of tokens to implement it can lead to compaction happening and then a lot of things can be lost in that compaction and then the final plan like won't be executed on properly. So what I personally did before this update was introduced is I would get it to write the plan to a brand new file and then load that file into brand new chat and then continue on and do the plan execute on the plan in that new chat. But now cloud code provides you an option to do that automatically. So if you press yes clear context and auto accept edit basically the final plan will just be loaded into brand new chat and then as you can see here we have way more room to work with and actually finish on executing the plan. But in this case, if I ignore clearing the context and do yes, then it will trigger the hooks that I have earlier. And yeah, you can see here when it finished planning, it took about 78,000 tokens, which means that I probably should have done it in the brand new context window. So you can see it's called the two hooks here, archive plan, review plan. And in archive plan, basically I now have it right over here. So this is the archived plan in a dated folder. And then I have some plan reviews as well. So this is a review that Codex CLI gave the plan. And then this is a review that Gemini CLI gave the plan. And now that goes back into cloud code. Say it says external review flagged several important issues. Let me address those issues. Good feedback from external reviewers. And now it's going to improve on the plan. And actually I forgot to mention we did not block the plan from happening because claude code is now going ahead and actually executing on the plan here. And that's because the verdict that uh Gemini CLI gave is basically warning about issues with the plan. And the verdict that Codex gave is a warning as well. If Codex and Gemini said like, "Hey, we're going to block the plan," then it would redo the plan basically. So, if you are interested in this hook and workflow to automatically review plans with other CLI tools, then you can either create the hook yourself or join my Claude code masterass where you can download it. And there's another related lesson as well where you can have different sub agents with different skills review plans before you continue with the plan.

Claude Code has now added planning mode to desktop application. So if you make sure local is selected over here, you can press plan and then basically come up with similar plans like you would do in the CLI version. So this probably means that you can have multiple planning agents running in parallel considering plans from different perspectives. And I hope they do add a planning mode to a cloud version because it can be quite handy to plan from your phone or have many different plans running on the cloud and then picking the best one. Now this one is pretty

interesting because you can autoenable MCP tool search if your MCP tools were to use more than some percentage of your context window. So basically what they did recently is they added the ability to search MCP tools because before in cloud code if you had a MCP server enabled then it would load every single MCP tool and the description of each of the tools. So previously it would look kind of messy where you had the name, description, input schema and a lot of other things about every MCP tool loaded into context at once. But now we have one tool called MCP search that is given the name of all the tools that you have available from MCP servers and then it can like fetch and get more information about each tool if needs be. Now this requires every MCP server to have a descriptive name for each tool that they have available. But essentially what it means is that by default if cla code sees that your MCP tool descriptions would take up more than 10% of a context window then it will automatically enable MCP tool search which can give you some token savings which in turn gives you better accuracy for longer as well. But now you may actually want to change that 10% to another number or always have it be on and you can do this pretty easily in your settings. json by adding this as an env and defining the number here. So 20 here would be 20%. So going to settings file and defining en we can set it to either auto which is 10% by default true false so auto then a number we specify. So I'm going to set it to auto 5 instead because I think 10 is too much. But yeah it's now pretty interesting because coding agents are now adopting the standard where things are being progressively disclosed to coding agent when needs be. So we already have skills. We now have like MCP search tool and I'm guessing we're going to have a lot more things like that in the future. Now, some people were using MCP portter previously where it would call the MCP tool with a CLI command. But it turns out the cloud code team tried both and found the MCP search tool to work better because of like the nuances of people's different environments. We now have a brand new

hook that I've covered in more detail in my master class. But essentially, if you define this hook in your settings folder, then you run cloud- in it, then it will run cloud normally, and it will also trigger that hook as well. So, you can have any commands that run automatically on each session of you running cloud code. If you want to do the same thing in headless mode instead, you can do in it only where it will basically run that hook but not actually spin a possession of cloud code. So that can be handy if you're using in CI/CD pipelines. Then we have cloud maintenance where it will run cloud code and also trigger another script that you may have. So this kind of looks like this. Basically the way you would differentiate between the three modes is you would make a hook dispatcher that looks kind of like this. Then you would trigger a different script each time. So you would have like a script that would automatically install any mpm packages on your computer. A maintenance script that would basically make sure that everything is like up to date for example. And you could have it do a bunch of other things like initialize a repo for a brand new person you're on boarding onto a project. Start services locally on your computer like a database for example. Or with maintenance you could do like formatting or updating documentation by spawning up more instances of cloud code and doing stuff like making sure dependencies up to date. So it seems they're trying to integrate it more into team workflows. So you can basically onboard someone tell them to run a claw d- init and then the project will be automatically set up on their computer. Pre-tool use hooks

now support additional context that can go to model. So previously in cloud code if cloud code was going to run a command to remove data for example then it might ask the user for permission with a reason that is shown to user but not to cloud code. But now we have additional context that can go back to cloud code to help refine its next steps if it tries to do something dangerous. So as an example, we could allow a tool to be called and then pass in additional context saying that we are in the production environment and that you should proceed with caution. The web and

desktop version of cloud code now includes a diff view where it shows the lines deleted and added and you can include your own suggestions and comments by clicking on any of the plus buttons and then typing something in here. So pressing enter on a suggestion, it will then send it as a brand new message with the line and file reference. When cloud code now asks for

your permission to do something, if you choose yes, then you can also press tab and then give it additional instructions by saying something like uh for example, I do wanted to check out this branch, but call it something else. So say call it 11 labs instead. Press enter and then that will be passed alongside the approval for that particular permission.

They added these two variables that allow the status line to have a better display of the remaining context. So you could copy this over and then go to cloud code do/status line to spot of the status line sub aent paste that into here and then say can you update the status line battery to included this instead for the current context window. It now updates the command so it doesn't have to do the calculation anymore. The

at autocomplete now added icons for different types of suggestions. So doing at and typing something plus basically refers to any files. star refers to any agents and then diamond refers to any MCP server tools. They now merge /

commands and skills together. So you can still have the commands folder for example, but behind the scenes in Claude Code, they are treated as the same

thing. And then finally, in the desktop app, they added the ability for you to move Claude the crab around by just like hovering over it. Anyways, I do go into

many of the things in much more detail in my Claude Code Masterass. There will be a link down below to check it out. And the new year sale will be ending soon. And over time, the prices will become more expensive as I continue to add more videos and features.