Curated tools, libraries and resources worth remembering. Organized by what you're trying to do — read the one file that fits the question, not the whole set.
| File | Reach for this when… |
|---|---|
| video.md | Making, cutting or understanding video |
| voice.md | Text-to-speech, speech-to-text, dictation, lip-sync |
| browser.md | Driving a browser, scraping, or end-to-end testing |
| ui-design.md | Getting the agent to produce UI that doesn't look generated |
| ui-libs.md | Component libraries, icons, editors, canvas, visual assets |
| code-quality.md | Review, language servers, understanding an unfamiliar codebase |
| security.md | Scanning your own code, reverse engineering, authorized pentest |
| prod-services.md | Connecting the agent to databases, payments, observability, infra |
| marketing.md | SEO, paid media, content production |
| agent-tooling.md | Building agent extensions, changing agent behavior, alternative runtimes |
| collections.md | Skill sets and indexes to install from |
| automation.md | Workflow automation and integrations |
| inbox.md | Untriaged — staging area before something earns a home |
Every entry carries a type marker, because a domain file mixes kinds and the marker tells you how you'll be installing it:
[MCP] server · [Plugin] · [Skill] · [Library] · [CLI] · [App] · [Collection] · [Reference]
The three install syntaxes, given once here rather than repeated per file:
- MCP servers —
claude mcp add <name> <endpoint-or-command>. Many are remote/OAuth, so first use opens a browser. - Plugins —
/plugin install <name>@<marketplace>, usually preceded by/plugin marketplace add <owner/repo>. Restart Claude Code after installing. - Skills —
npx skills add <owner/repo> --skill <name>. Symlinked by default sogit pullkeeps them current;--copyfreezes a snapshot. See vercel-labs/skills.
Only add what fills a gap your workflow doesn't already cover. Brainstorming, spec-writing, planning, execution, research, explanation, capture, debugging and verification are already owned. A tool that duplicates one of those competes with it rather than helping.
Treat every MCP server, skill and plugin as executable code you're granting access to. Skim the source, scope credentials tightly, and prefer read-only where the option exists. Treat any skill-discovery tool as untrusted input.
A mid-2026 benchmark found most SWE skills produced no measurable gain, and some hurt when they conflicted with project context. Star counts measure attention, not usefulness. Where two tools overlap, the files here say so and pick one — follow that rather than stacking both.
Domain first. A domain earns its own file at 3+ entries; below that it's a section in the closest existing file. When you're unsure, it goes in inbox.md until it earns a home.
Snapshot: mid-2026. This ecosystem churns fast — check a repo's recent activity before installing, especially community entries. Entries known to be stale carry a dated warning.