All five tools are real and all five work. But type the circulating list as-is and two of them install the wrong package; in two of these projects the repo name and the package name don't match. Two of them now recommend a CLI install over MCP, and the reason is worth knowing on its own. I verified everything against its source.
Claude CodeEklentiGeliştirici Araçları
The repo name is not the package name
Every link in the list points at the right repo. The problem isn't the links, it's seeing the repo name and assuming it's the package name. In two of these tools they differ, and in both cases the wrong guess quietly takes you somewhere else.
✗npm install -g playwright-cli. That name really does exist on npm, so it won't error. But the package's own description says: deprecated, use @playwright/cli instead.
✓npm install -g @playwright/cli@latest. It still installs a command called playwright-cli; only the package name changed.
✗npx npxskillui. The repo is called npxskillui, so this is everyone's first try. No such package exists on npm.
✓npm install -g skillui. Both the package and the command are just skillui.
Install as MCP, or as CLI?
Two tools on this list, Playwright and Context7, no longer present MCP as the only route; they point you at the CLI. The reasoning is the same in both: the moment an MCP server connects, it loads all of its tool definitions into the model's context. Part of your window is spent before you've asked anything.
On the Playwright side the gap is wider still, because the accessibility tree coming back from the browser also lands in context. The project's own docs say it outright: CLI invocations are more token-efficient. Same work, less of the agent's memory spent on it.
Installing as MCP
the model sees the tools up front
The agent sees the tools listed; it doesn't need to read anything else
Every tool schema loads into context the moment it connects
Stack a few servers side by side and the window fills fast
Installing as CLI + skill
the model learns it when needed
Only a short instruction file is loaded
The agent reads command details from --help when it needs them
Does the same job on noticeably less context
Two things before you run Strix
The other four run as soon as they're installed. Strix is different: it wants Docker running and your own model key. The tool itself is open source and free, but every scan you run lands on your API bill, a part these lists never mention.
Only point it at your own systems
Strix genuinely attacks the target: it doesn't just find a hole, it exploits it to prove the finding is real. The repo's own warning is blunt: only run it against systems you own or have explicit, written permission to test. Unauthorised scanning is a crime in most jurisdictions. On your own project it's fine; on someone else's site, good intentions are not a defence.
01
Playwright CLI: giving your agent a browser
Lets the AI actually open and look at the page it just built: clicking, filling forms, taking screenshots, reading console errors. That's how it catches its own mistakes before you see them. Two commands; the second installs the skill that teaches Claude Code how to drive it.
The package is @playwright/cli; the repo is playwright-cli. Install the bare name and you get the deprecated old package. After setup, playwright-cli --help lists the commands, and that's where the agent reads them from too. To watch the browser while it works there's a dashboard: playwright-cli show. The package is still on 0.x, so commands may shift.
Supabase plugin: running your database from the terminal
Lets you build and manage your app's database from inside Claude: creating tables, writing migrations, deploying edge functions. One command installs both the MCP server and two skills: one for Supabase generally, one for Postgres schema, query and row-level-security patterns.
The command finds your installed agents by itself, Claude Code included. The server signs you in through the browser; you don't need to create a personal access token by hand. If you don't want the agent holding write access to a live database, append read_only=true to the server URL, or project_ref=<project-id> to lock it to a single project. Don't skip those two on production.
Vibe-coded code leaves holes behind very easily. Strix pokes at the app like a real attacker: it finds the hole, confirms it with a working proof, then proposes the fix. That's the difference: not a list of guesses, but results it actually tried.
Docker has to be running; the first run pulls the sandbox image. OpenAI isn't required; Anthropic and Google work too, and you write it as provider/model on the STRIX_LLM line. Results land in a strix_runs folder. The install line runs a script straight off the internet; if you want to see what it does, open that URL in a browser and read it first.
SkillUI: extracting the design system of a site you like
Crawls the site you point it at and pulls apart its design system: colours, fonts, spacing, animations, component patterns. It writes the output into a folder Claude can read, and you build your project in that language. Nothing goes to the cloud and no AI is involved; plain analysis, entirely on your machine.
Installation
npm install -g skillui
skillui --url https://linear.app
cd linear-design && claude
The repo is npxskillui but the package is just skillui. The default mode is static analysis; for screenshots and animations add --mode ultra; that mode also needs npm install playwright and npx playwright install chromium. Set how many pages to crawl with --screens; default 5, max 20. Requires Node 18+. Use it to learn a system and adapt it to your own work; the licences on font files and brand assets don't transfer to you.
What the model knows freezes on a date; libraries don't. Context7 pulls the current docs for the package you're using and feeds them to the model as you write. That gap is most of why you write code against a stale API and then debug it for hours.
Installation
npx ctx7 setup --claude
Setup asks which of two modes you want: skill mode installs a file that teaches the agent to use the ctx7 command, MCP mode registers the server. Skill mode eats less context and is the recommended route. The old MCP command still works, but it's no longer the only way. A key isn't required; grab a free one at context7.com/dashboard for higher rate limits.
Installing these tools on your own is one thing; actually building with them is another. The community has people using these daily and people building systems from scratch.
DOA: Yapay Zeka ve Otomasyon
If you want a system that actually runs in your business, let's talk for 10 minutes; I'll look at what you're trying to build and tell you which path fits. Free, and not a sales pitch.