Make Claude Code 10x More Powerful (Only Plugins You Need)

0:00 Open Claude Code, type /plugin, and you land in a catalog holding 286 plugins. You did not add that catalog. Anthropic's own marketplace installs itself the first time you start the

0:11 tool. Finding plugins is the easy part. Working out which handful what they cost you to keep is the whole job because none of them are free. Anything you leave switched on gets build again on

0:21 every single turn you take. Anthropic publishes a table of what each kind of extension costs you. six rows and one of them is strange. Five of those rows describe rent. The sixth says the cost

0:33 goes down. It is the only line like it on the page. That row is called code intelligence and on most of the machines watching this, it is switched off right now, not missing off. The tool behind it

0:44 has been sitting inside Cloud Code since the 19th of December. It ships in the binary. It does absolutely nothing. And it waits there for one small plugin to wake it up. There are 12 of those

0:54 plugins, one for each language. and you will only ever install one of them. So, what does a real language server hand a coding agent that searching through your own files cannot? Start with the cost

1:04 because it is the part most people skip. There is a page in the claude code documentation that puts a price on every feature, six of them, each with a straight answer about what it takes out

1:14 of your window. Your claude.md loads at session start full content and the cost column says three words every request. Skills load their descriptions at the start and the full text only when used.

1:26 MCP servers load tool names and hold the schemas back. Hooks are the cheapest thing on the list because they load nothing at all. They run outside the conversation, so their cost column is a

1:37 single word, zero, unless the hook hands something back. Every other row is a subscription you pay per turn. Then the code intelligence row, it loads after file edits and on demand. And where

1:49 every other row names a price, that one reads low, reduces file reads elsewhere. Five rows are rent. That one is a refund and it is the only line like it. The mechanism is dull enough to be easy to

2:00 miss. Looking a symbol up returns a file and a line number while reading the file to find that symbol returns the whole file. So the extension that makes the model sharper also makes your context

2:10 smaller. Those two things are normally opposites. But cheap is not the same as good. and a table in a document is not evidence that a thing works. Therefore, the tool underneath that row and the

2:21 date that it arrived on. Claude code shipped a language server tool in version 2.0.74 published on the 19th of December 2025. The change log line is one sentence.

2:33 Added LSP tool for code intelligence features like go-to definition and find references that landed 71 days after the plug-in system itself launched. The plugins arrived first in October and the

2:44 capability they exist to switch on turned up 10 weeks later in a point release on a Friday with no announcement of its own. It has been in the product for 8 months and the documentation

2:54 carries the note that explains everything else. The LSP tool is inactive until you install a code intelligence plugin for your language. Inactive is their word, not mine. The

3:04 thing never starts, which is why it is easy to have never seen it work. There is no error message, no warning, no missing menu item. It simply does not turn on. And an agent that has not had

3:15 it will not tell you because it has nothing to compare itself against. Without it, Claude finds things the way you would with no editor open. It searches text. GP across the repository,

3:26 read whichever files matched, reason about what came back. On a small project, you would not notice the difference. On a large one, you notice constantly because a text search cannot

3:35 tell a definition from a mention or a real call site from the same word sitting inside a comment. With a language server, Claude gets two capabilities instead. And the first is

3:45 underrated diagnostics. After every single edit Claude makes, the language server reports the error straight back to it. Anthropic's wording is worth quoting exactly. If Claude introduces an

3:56 error, it notices and fixes it in the same turn, not the next turn, the same one. The second is navigation. Jump to a definition, find every reference, read a type on hover, list the symbols in a

4:07 file, trace a call hierarchy. Those are the operations your editor has had for a decade and the ones a text search can only approximate. The documentation is careful about that claim, and so should

4:18 we be. Their phrase is more precise navigation than GRP based search. And then in the very same sentence, a hedge, though availability may vary by language and environment. A vendor writing

4:29 carefully. 12 plugins cover 13 languages because clangd handles C and C++ together. Python, TypeScript, Go, Rust, Java, C, Cotlin, Swift, PHP, Lua, Ruby, and the C pair. You install the one for

4:44 the language you are actually working in. And here is where the popular version of this tip is wrong, including the version handed to me. It is not one line and it does not work instantly. The

4:55 documentation says so directly in the sentence immediately above that table which is the sentence people skip. Install the language server binary from the table below before using these

5:04 plugins. The plugin does not install it for you. So Python wants Pyrite on your path. Rust wants Rust analyzer. Go wants Go pls. The plugin is the wiring. The binary is the engine. It is still one

5:16 install command with a package manager you already have. But one command is not instant and a tip promising instant gets abandoned at the first error. That error reads executable not found in path in a

5:28 tab of the plug-in panel few people open. There is one more wrinkle and I only found it by reading the catalog file rather than the documentation. The documentation table lists 11 languages.

5:38 Run down it alphabetically and it goes PHP, Python, Rust. Ruby is not there at all, which would be unremarkable except that the catalog cla code actually reads has a Ruby plugin in it authored by

5:50 Anthropic sitting exactly where the table skips over it. The repository has a folder for it that matches the other 11 exactly. Same license, same shape, same kind of readme naming the binary.

6:02 So, Ruby is installable and undocumented. Nothing here is broken. It is simply missing from the one page you would check first. The catalog is the truth and the documentation is a summary

6:12 of it. That same documentation carries the counterargument which I respect them for printing. Language servers like Rust Analyzer and Pyite can eat memory on large projects and inside a monorreo

6:23 they can flag imports that are not really broken. The advice then is blunt. Disable the plugin and use Claude's built-in search. That is a vendor telling you when its own recommendation

6:32 stops being right. Now, the title on this video promises 10 times more powerful, and I owe you something straight about that number. No controlled measurement of what code

6:41 intelligence does to an agent's accuracy has been published. Not by Anthropic, and not by anyone I could find. What is measured is narrower, and I think it is the better signal anyway. 25 separate

6:52 change log entries touching this one subsystem. Memory leaks, zombie servers after a crash, file paths on Windows, a command injection bug in the code that finds the binary. That is not marketing.

7:03 That is eight months of maintenance. So, the biggest single upgrade, yes, 10 times unmeasured and I would rather say so than pretend, which leaves 274 other plugins to account for. The catalog

7:15 answers that faster than any review could. Of the 286, Anthropic wrote 38. The other 248 are somebody else's code that Anthropic has agreed to list. And the docs are blunt about what listing

7:28 means. They cannot verify that plugins work as intended. So, of the 38 they did write, which did they care about most? 12 are the language servers. Roughly a third of everything the company shipped

7:38 into its own marketplace is one idea repeated per language. A strong opinion and a strange place to find one, a directory listing rather than a blog post. Of what remains, here is the short

7:50 list. The first is what I would install before anything else, and it is not exciting, which is probably why it gets skipped. It is called claude code setup and its best feature is that it changes

8:00 nothing. The catalog describes it as analysing code bases and recommending tailored automations, hooks, skills, MCP servers, sub aents. It is read only so you make it argue rather than guess. It

8:13 was written by Isabella he at anthropic who also wrote the second one and the second is about the file you are worst at maintaining. claude.md management audits that file against your actual

8:24 codebase and captures what a session taught you at the end of it. That matters because the documentation recommends keeping claude.md under 200 lines and every one of those lines is

8:33 built on every request forever. A claude.md left untended for 6 months is the most expensive file you leave unopened. Third, context 7 from Upstash. It fixes the complaint I hear about

8:46 coding models more than any other. And its own page states the problem in three blunt lines. Code examples outdated and based on year old training data. Hallucinated APIs that do not exist.

8:57 Generic answers for old package versions. Context 7 is a hosted M. CP server that pulls version specific documentation straight out of the source repository into the prompt. Nothing to

9:08 install. It works without an account and a free key lifts the rate limit. Fourth, Hookify. And this one changed how I think about correcting an agent. The reason is a single line in the anthropic

9:19 docs. An instruction like neveredit.env written into your claw.md is a request, not a guarantee. A hook that blocks the edit is enforcement. Hookify closes that gap because it writes the hook for you

9:31 from a sentence. You type /hookify warn me when I use rm-rf and it writes the rule file. It takes effect on the very next tool use. and hooks are the row in the cost table that reads zero. Fifth

9:44 and sixth are the two security plugins and the useful thing to know is that they are not the same tool twice. Security guidance runs while Claude is writing. Pattern warnings fire on about

9:54 25 known dangerous patterns as the edits happen. Then when Claude finishes a turn, it sends the diff to a fast model opus 4.7 by default and feeds the serious findings back to be fixed before

10:06 you see the reply. Then on commit, an agent reads the surrounding files to trace data across them. That last layer catches the bugs a pattern match cannot see. It was written by David Dorin, a

10:17 security engineer at Anthropic, whose address sits in the plug-in's own manifest. The other one, Claude Security, is the deep scan and its own description is what sold me. Every

10:26 finding is challenged before it is reported, and the verification tally is computed in code rather than asserted by a model. A scanner that argues with itself and then counts the argument is

10:36 rarer than it ought to be. But read its trust model before you point it anywhere. It runs inside your session under your permissions and it adds no isolation of its own. Their readme says

10:47 it fits code you control where the question is which bugs are in the code rather than whether the code is trying something. That is a real distinction. The rest of the set is quick because

10:57 these are tools rather than arguments. Pull request review toolkit gives you six specialists instead of one generalist. Skill creator builds your own skills and runs evaluations against

11:06 them. Code simplifier cleans up what you just changed rather than the whole repository. Plug-in dev and MCP server dev are for when you start adding to this ecosystem instead of consuming it.

11:17 The integrations are what they sound like. GitHub and GitLab for repositories and pipelines. Playright from Microsoft for driving a real browser. GPile for asking your repository questions in

11:28 plain English and Terraform which is Hashi Corp's own server rather than a strangers which brings me to the entry that complicates this whole video and I would rather show it to you than

11:38 summarize it. Serena is an Anthropic official marketplace. It does semantic code navigation. It wraps language servers for over 40 languages against Anthropics 12 and it pairs with

11:49 everything I have just recommended and its own readme opens with this. Do not install Serena via an MCP or plug-in marketplace. They contain outdated and suboptimal installation commands.

12:00 Instead, follow our quick start instructions. That is the project telling you not to take it from the shelf that it is sitting on. Both things are true here. The listing is real and

12:09 the tool is good. It just means that official marketplace has meant listed and it has not meant maintained. Anthropic says as much itself. They cannot verify that plugins work as

12:19 intended. Outside that catalog, five community repositories are worth your time, and the healthiest belongs to a name, Anthropic, printed in its own launch post. Seth Hopson had been

12:30 collecting specialist agents since July of 2025, 11 weeks before plugins existed. On launch day, Anthropic pointed at him for curating over 80 sub agents. Today, that repository holds 92

12:42 plugins, 202 agents, and 181 skills. And it runs on cursor, open code, anti-gravity, and co-pilot, as well as clawed code. 39,000 stars, and four open issues, which is the best ratio of

12:56 anything I looked at for this video. Superpowers was created on the 9th of October, the same day the plug-in system shipped, and now carries 275,000 stars. Matt PCO's skills repository has

13:08 230,000. Planning with files solves one problem. Well, an agent forgetting its plan after a compaction. and Claude Hud by Jared Watts updated four days ago puts your

13:20 context usage on the status line which turns everything I said about cost into a number you can watch tick the plugin that gets screenshotted most is the best story in this ecosystem and the weakest

13:30 recommendation in it openai shipped a plugin for claude code not a community port the openai organization an official plugin whose whole job is to delegate work to open aai's codeex from inside

13:41 Anthropics tool 32,000 stars and it works two rival labs, one terminal. But it was last pushed on the 8th of July, 45 days ago, 6 and 12 weeks in an ecosystem where the healthy repositories

13:55 shipped this week. And its issue count is where I have to correct a number that nearly every article quotes, including the version of it that was handed to me. People repeat 414 open issues, open the

14:06 repository page, and GitHub splits it for you. 224 issues, 190 pull requests. The programming interface adds them into one field and every tool inherits that. The correction runs the other way too.

14:20 Seth Hopson's 9 is four issues and five pull requests. If you have ever judged a repository by that number, and I have, you are judging a sum of two different things. So, the verdict, and I am not

14:31 going to hedge it, install the language server plugin for your language before anything else. For anyone in a typed or compiled language or a codebase big enough that Grep has started to feel

14:41 slow, it is the largest upgrade on the shelf and is the cheapest one on it. The 5% it is wrong for are real. Anthropic names them rather than me. A language with no server in the set. A monor repo

14:53 that makes the server cry wolf. A machine where Rust analyzer eats the memory. There their own documentation says switch it off. Then take four or five more and stop. Cloud code setup to

15:04 tell you what your repository wants. Context 7 so the model stops writing last year's API hookify so your rules are enforced instead of requested. One security reviewer and a status line

15:16 because the thing worth remembering is what Anthropic wrote on the day it launched all of this. They are designed to toggle on and off as needed. Enable them when you need specific capabilities

15:26 and disable them when you do not to reduce context and complexity. That was the instruction from day one and the ecosystem read it as an app store where more installed is better. Anthropic has

15:36 since had to ship a feature listing the plugins you installed and have not touched in 2 weeks under a heading reading not used recently. You do not build that feature unless people are

15:45 hoarding. So the open question is not which plugins to install. It is this. If the most valuable extension in the whole catalog turned out to be the one that ship switched off, what else is sitting

15:56 inside the tools you use every day?