Claude Mods - The Biggest Claude Code Upgrade!

Author: Prompt Engineering · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Claude Mods - The Biggest Claude Code Upgrade! · Channel: Prompt Engineering · 10:30 · Watch on YouTube

Topics: AI agents, Cybersecurity and software security, Claude Code, AI developer tools

Summary

Claude Code mods run inside the agent harness, unlike hooks and other extensions that operate externally, and can change or stop actions, affect what appears on screen, and add persistent session features. Mods can be loaded from plug-ins and hot-reload when their files are saved, though a reload resets in-memory state. The video demonstrates custom panels, command guards, tool-use counters, and a context-window gauge, and shows how Claude can suggest mods based on a user’s conversation history. Because mods can access files, run programs, and use the network, the speaker advises validating and carefully reviewing them before installation.

Key points

Mods run inside Claude Code and can intervene in the agent harness, rather than being called from outside it.
A mod can let an action proceed, alter it before continuing, or stop it and provide a response to Claude Code.
Mods can affect what users see, such as hiding email addresses in displayed responses while the model still sees the original address.
Unlike one-off hooks, mods stay loaded for the session and can retain state, display live panels, ask for confirmation, and add slash commands.
Mods are installed through plug-ins; saving a mod file reloads it in the current session, but resets its in-memory state.
Examples include a tool-use counter, a local-model status panel, a command guard, and a context-window gauge.
The speaker suggests asking Claude to review the last 20 sessions for recurring questions, concerning commands, and ways the user checks its work, then present ideas before building anything.
Mods can read files, run programs, and use the network, so the speaker recommends validating unfamiliar mods and reviewing their events and calls.
Use hooks for simple script-based blocking, skills for instructions, MCP servers to reach other services, and mods for interactive or persistent changes to the Claude Code experience.

Chapters

Start	Chapter	Summary
00:00	What mods are and how they work	The speaker explains how mods differ from hooks and other extensions, and how they can intercept commands or change what appears on screen.
03:40	Why mods matter and getting started	The video contrasts persistent mods with hooks, describes Claude Code features built as mods, and shows how to load and try one.
06:21	Examples and personalized mods	The speaker demonstrates a status panel and command guard, then describes generating useful mods from a user’s own session history.
08:55	Security and when to use mods	The video warns that mods can access files, run programs, and use the network, and explains when to choose mods over hooks, skills, or MCP servers.

Quotes

Mods on the contrary live inside cloud code itself — 00:31

A mod gets loaded once and stays running for the whole session. — 03:40

Show me the ideas first. Don’t build anything until I pick. — 07:52

don’t just blindly trust the mods that you find on the internet and install them. — 09:25

Concepts and entities

Claude Code, Anthropic, mod, hook, MCP server, skill, DeepSeek, Claude

Full transcript (original language)

00:00 Okay, so Enthropic just dropped mods in cloud code and this is a much bigger deal than a lot of people realize. This is the first time Enthropic has really opened up clot code itself and you can now customize it in pretty much any way you can imagine including the UI itself. You can basically create your own custom harness on top of cloud code which is pretty exciting. Now mods are very different from things that we have had before. Things like scales, hooks and

00:31 MCP servers. All of those live outside cloud code and cloud code calls them when it needs them. Mods on the contrary live inside cloud code itself, which means they can really control the behavior of the agent harness. So they can essentially step in right before a cloud code takes an action and change it or stop it. Now the idea is not new. We have seen it with DeepSeek. When Deepseek released their harness, uh they

01:02 said everything in it is a plug-in. Things like the model wrapper, the tools that you can use, even the agent loop itself. And the nice thing is if you don’t like one of those parts, you can just swap it out with your own. Now, Entropic is bringing a very similar idea to cloud code with mods. So, in this video, I want to explain what mods actually are, how to install them, and then a really easy way to automatically create mods from your own conversation

01:33 history. That way, those mods are going to be actually useful to your own workflows. First, let’s talk about what exactly is a mod. Technically, a mod shapes as a plug-in with some extra settings. So it has the same plug-in file, same hook file except that the hooks file now points to a JavaScript or TypeScript file and this new file is your mod which runs inside cloud code and there is no extra build step. So when you load that

02:05 one flag, cloud code watches that specific folder. So every time you save the file, the mod reloads in the session you already have open. Now, how does a mod actually work? So, let’s take an example. Say Claude is about to run a force push. Before it does, cloud code calls your function and it hands it three different things. The first one is the event itself, which is what’s about to happen. In this case, the exact command. The second is a way to talk

02:37 back to clot. So if you have some interesting internal steps that the mod has to take things like maybe draw on the screen or read a file or ask you a question those are going to be a part of the second thing. And the third is a function called next which basically means carry on as normal. So your function has three choices. It can call next and let the push go through. Maybe it’s going to log it along the way. Or it can change the command first, say

03:09 push to a new branch instead, and then call next or it can skip next completely and answer it for clawed code. Things like no, we don’t force push in this specific repo. Now, the thing is that it’s not only for commands. A mod can sit in front of what claude shows you. For example, you don’t want to show any PPI or personally identifiable information like emails in the responses. So then cloud using the mod can hide email addresses in the

03:40 response. Now the model still sees the real address. It just changes the behavior on the screen. Okay. Now you might be thinking we have hooks for things like this. So why is this such a big deal when we already have had hooks? So a hook is a shell script. It runs once for an event, send backs the answers, and then it’s basically gone. A mod gets loaded once and stays running for the whole session. So it can remember things. It can show you a live

04:12 panel that updates every turn. It can pause commands and ask you before it runs. It can also add slash commands that runs your code instantly without calling the model. So it’s a lot more than a hook. And here’s the part I actually find interesting for this implementation. Entropic is already building parts of the cloud code itself as mods. For example, the diff view and the agent. FD supports are both mods. And they have said that they planned to move more features over as mods. So this

04:46 gives you the ability to really customize the whole clot code experience for you. Your clot code may not look similar to mine anymore. Now it’s the same core uh with different set of mods on top and that’s what I mean when I say anthropic has opened up cloud code until now. You could add things around it. Now you can change the thing itself. Okay. So let me show you this running. So the first thing is you need to update your cloud code instance or your cloud

05:18 desktop app to get access to it. If you are on the latest version, you already have mods. So if you type /plugin and go to the install tab, you will see a few built-in ones already there. And one of them is the skill Claude uses to write mods for you. Okay, so here’s the simplest mod from the docs. It counts how many tools Claude uses. Now I loaded this with the uh plug-in directory flag and ask cloth to read a couple of files

05:48 for me. So it shows the spinner. It now shows the tool count. In this case, the code was drawing inside the clot code. Now if I type /tally, it actually shows three. I also edited the file while the session is still open and clot code reloads it on its own. But then the tally of the tool calls went to zero. That’s because a reload restarts your code fresh. So everything you want to keep has to be stored in clot code, not in your file. So installing uh someone’s

06:21 else uh’s mod works like any other plug-in. In this case, a marketplace is just a folder or a GitHub repo with one JSON file in it. There are two different install commands. And then /plugin uh now says two mods are active. The bar above the prompt that you see is the one I built for my uh lab. It shows which local models are running on my MacBook and on my two DJX Sparks. Right now, the

06:51 DJX Sparks are off and it tells me that. So, it’s just a neat little mod that can actually keep track of my DJX Spark. Now, here’s another one. Uh, this one is a guard. So I asked clot to delete a folder with rm-f before the command runs. The mod stops it and asks me I can say refuse or cloud code reads my reason and suggest a safer way and and when the turn ends a little pop-up tells me how long it took and how

07:21 many tools it used. Now in this case that single mod is doing three different things. Now the beauty is that you don’t have to write these yourself. I just described what I wanted. for example, a gauge that shows how full my context window is. Claude wrote it asked me once if I want to turn it on live on reloading and the gauge basically showed up above the prompt. And as I keep working, it’s going to basically show me how much my contacts window is filled up

07:52 till now. Now, these were just a few examples, but the most useful mods are the ones built around how you work. Now, here’s a prompt that I use for it. It tells clot to read your last 20 sessions and look for three things. The question you keep asking, the commands that make you nervous and how you check its work, and the most important line is the last one. Show me the ideas first. Don’t build anything until I pick. Now, I found this on X. I’ll put a link to the actual poster. Now, I ran this on my own

08:24 sessions, picked the ones I liked, and then Claude built those for me. So, this is my Claude code. You know, a panel that shows both Sparks and my Mac. Then a shelf of everything this session produced. A guard that checks which account I am about to publish from. Now none of these came from a template. These are based on my own history of using cloth and essentially I am creating my own version of cloth code which is pretty amazing. Now, one

08:55 warning, a mods runs on your machine. So, it can read your files, run programs, and use the network. So, before you install anyone’s else’s mod, run cloud plug-in, validate on it, it lists every event that the mod listens to, and every call it makes. So, for example, my lab mod makes network calls, which makes sense. My guard asks me questions and this community mod only hides my secrets from claude. So if a

09:25 mod you don’t know is making network calls, you want to know why, right? So very similar to skills, the MCP servers, don’t just blindly trust the mods that you find on the internet and install them. These are big security risks. So when should you use a mod? If you just want to block something with a script you already have, I think a hook is still fine. If you want clot to follow instructions, for that you want to use a skill. And if you if claude needs to reach another service, you want

09:57 to use an MCP server for that. But if you want something on your own screen, something that remembers or something that can stop cloud and ask you first, or you really want to customize the whole experience of cloud code, this is where you want to use a mod. Now, my guess is that we will see a lot of mod marketplaces over the next few weeks and soon after that, the first mod that does something it shouldn’t. Now, let me know uh in the comments what you would change first about your cloud code if you are able to use a mod. The prompts and the

10:30 mods are linked below. Uh anyways, I hope you found this video useful. Thanks for watching and as always, see you in the next one.