10 Claude Code Skills, Plugins & CLIs to Install on Day One (April 2026)

Skills vs Plugins vs CLIs (What the Labels Actually Mean)

Before we get to the list, a quick taxonomy. The three categories get conflated constantly because they all "extend Claude Code," but they do it in three very different ways and knowing the difference makes it much easier to pick the right tool for a given job.

Skills are markdown files. A skill lives in ~/.claude/skills/YourSkill/SKILL.md and contains instructions, prompts, reference material, and optional helper files. Claude loads a skill on demand — you (or Claude itself) invoke it by name, the contents get injected into the context, and then Claude uses them as guidance for whatever it's about to do. Skills are how you teach Claude your workflow.

Plugins are MCP (Model Context Protocol) servers. A plugin is an executable that starts up when Claude does and exposes one or more tools that Claude can call — like Read, Write, or Bash, but custom. Firecrawl gives Claude a scrape tool; Playwright gives it a browser_click tool. Plugins are how you give Claude new hands.

CLIs are independent command-line tools. They don't know anything about Claude — they're just programs on your PATH that Claude can shell out to using its built-in Bash tool. Codex CLI is a standalone OpenAI agent; Claude invokes it the same way it invokes git or npm. CLIs are how you give Claude helpers.

1. Codex CLI — OpenAI's Agent as a Sub-Agent

Counterintuitive day-one pick: install OpenAI's Codex CLI, a direct competitor to Claude Code, and let Claude Code invoke it as a helper. Codex is excellent at executing well-scoped implementation tasks — "rewrite this file to use React Server Components" — while Claude Code is excellent at planning, coordinating, and holding multi-file context. Run them together and you get two frontier models cooperating on the same project.

npm install -g @openai/codex
export OPENAI_API_KEY=sk-YOUR_KEY

# Inside Claude Code, tell it what's available:
# "If you need a second opinion or want to parallelize a task,
#  you can shell out to codex via Bash with: codex exec 'task description'"

2. Obsidian — A Second Brain for Claude

Obsidian is a local markdown-based knowledge tool, and with the community MCP server it becomes Claude's long-term memory. Point Claude at your Obsidian vault and it can read, write, and link notes during a conversation — perfect for research projects, reading lists, client docs, and anything else you don't want to re-explain every session. The underlying files are just markdown, so you can commit your whole vault to git and treat it as part of your project.

claude mcp add obsidian \
  --command "npx" \
  --args '["-y","@obsidian/mcp-server","--vault","/path/to/your/vault"]'

3. Autoresearch — Karpathy's Autonomous Experiment Runner

Andrej Karpathy's autoresearch repo packages a small language model training loop and gives it to an agent so the agent can autonomously run experiments, track results, and propose architectural changes. Install it as a CLI that Claude Code can invoke, and Claude becomes capable of actually doing ML research — not just talking about it. If you've ever wanted an agent that can fine-tune a small model overnight while you sleep, this is the starting point.

git clone https://github.com/karpathy/autoresearch ~/tools/autoresearch
cd ~/tools/autoresearch && pip install -e .
# Claude can now invoke: autoresearch run --config experiments/gpt_tiny.yaml

4. awesome-design-md — UI Design Reference

VoltAgent's awesome-design-md is a curated collection of design-system markdown files — color tokens, typography scales, spacing primitives, component patterns — all formatted specifically to be dropped into an AI agent's context. Install it as a skill and Claude Code starts writing UI code that actually looks intentional instead of generic. It's the single biggest quality-of-life upgrade for anyone using Claude to build frontends.

cd ~/.claude/skills
git clone https://github.com/VoltAgent/awesome-design-md
# Claude will auto-discover the SKILL.md files on next launch

5. Firecrawl — Scrape Any Site to Markdown

Firecrawl is the web-scraping backbone for AI agents. Point it at any URL and it returns clean markdown, structured data, or a full site crawl. Claude Code with Firecrawl installed can pull documentation, read competitor sites, fetch API references, or ingest an entire blog into context — all via one tool call. There's a generous free tier and the self-hosted version is open source.

claude mcp add firecrawl \
  --command "npx" \
  --args '["-y","firecrawl-mcp"]' \
  --env "FIRECRAWL_API_KEY=fc-YOUR_KEY"

6. Playwright — Browser Automation MCP

Microsoft's Playwright MCP server turns Claude into a full browser driver. It can navigate pages, click elements, fill forms, take screenshots, capture network requests, and validate outcomes — all from natural language instructions. This is the plugin to install if you want Claude to actually use the web instead of just read it: filing bug reports, running end-to-end tests, scraping JavaScript-heavy pages, or validating your own app by clicking through it.

claude mcp add playwright \
  --command "npx" \
  --args '["-y","@playwright/mcp"]'

7. NotebookLM CLI — Google Research in Your Terminal

NotebookLM is Google's research assistant — upload PDFs, websites, and YouTube videos, get grounded summaries and cited Q&A. The community CLI wrapper exposes it as a terminal tool Claude Code can invoke. For any project where you're reading large volumes of source material and need grounded answers with citations, letting Claude hand queries off to NotebookLM is dramatically more reliable than asking Claude to remember everything.

pip install notebooklm-cli
notebooklm auth login
# Now Claude can run: notebooklm query "what does this PDF say about X"

8. Skill Creator — Anthropic's Official Scaffolder

The anthropics/skills repo contains a skill called skill-creator that writes new skills for you. You describe what you want — "a skill that runs our deployment checklist" — and it produces a properly-structured SKILL.md with front matter, description, instructions, and optional helper files. This is the fastest way to start producing your own custom skills, and because it's maintained by Anthropic, what it generates matches the official conventions exactly.

cd ~/.claude/skills
git clone https://github.com/anthropics/skills
# Inside Claude Code, invoke it: "use the skill-creator to make a new skill
# called deploy-check that..."

9. RAG-Anything — Multimodal Knowledge Graphs

HKUDS's RAG-Anything is a multimodal retrieval-augmented generation framework that ingests PDFs, images, tables, and text into a single knowledge graph. Install it as a CLI and Claude can query your private document collection with grounded retrieval — the results come back with citations and Claude can follow them to the source. It's the right tool for projects where you have hundreds of technical documents and need the agent to actually find the right paragraph instead of hallucinating.

pip install rag-anything
rag-anything ingest ~/docs --index ~/.rag/index
# Claude queries it via: rag-anything query "your question" --top-k 5

10. Google Workspace CLI — Full Gmail & Docs Control

The Google Workspace CLI gives Claude Code keys to Gmail, Docs, Sheets, Drive, and Calendar. Install it, authenticate once, and Claude can read your inbox, draft emails, create and edit documents, schedule meetings, and pull cells from spreadsheets — all from the terminal. For anyone running a Claude Code–centric workflow, this is how the agent leaves the developer sandbox and starts touching real work output.

npm install -g @googleworkspace/cli
gws auth login
# Claude can now run: gws gmail search "invoice" --max 10
#                     gws docs create "Meeting notes 2026-04-11"

Final Word

The reason Claude Code has pulled so far ahead of every other agentic coding tool in 2026 isn't the base model — it's the extensibility layer. Skills let you teach it, plugins let you extend it, CLIs let you compose it with every other command-line tool on Earth. The 10 above are our pick for the day-one stack because each one solves a distinct category of friction: second AI, knowledge, experiments, design, web, browser, research, scaffolding, RAG, and productivity integrations.

Install the four we flagged as essential first — Codex CLI, Firecrawl, Playwright, and Skill Creator — and play with them for a day before adding the rest. By the end of your first week, you'll know which of the remaining six match your workflow and which you can skip. The whole stack is free, the whole stack is open source, and the whole stack composes with everything else you already use.

For the full directory, browse our Claude Code Skills list, the complete MCP Servers catalog, our Claude Code Hooks guide, and the latest Claude Code Agents roundup. If you're comparing Claude Code against other tools, see the best AI coding tools of 2026 and our tier list of 24 Claude Code backends.

FAQ