Skilloadout MCP server
Commit Nov 13, 2026Provides a searchable catalog of over 12,000 safety-scanned AI agent skills for Claude Code, Cursor, and other MCP-compatible clients.
Also known as RaveenDi/Skilloadout
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is Skilloadout MCP?
Skilloadout provides a searchable catalog of over 12,000 safety-scanned AI agent skills for Claude Code, Cursor, and other MCP-compatible clients.
Summary generated from the project's README
About Skilloadout MCP server
Finding and installing specialized capabilities for AI agents is what Skilloadout enables for developers using Claude Code, Cursor, or Gemini CLI. This MCP server provides access to a massive catalog of over 12,000 skills that have been scanned for safety and checked for licensing compliance. It functions as a local plugin marketplace that requires no external accounts or backend hosting, as the repository itself serves as the product.
Users can interact with the catalog through specific tools like search_skills, get_skill, install_skill, and list_categories. The system is designed to be highly secure, using a two-pass process that includes automated scans for destructive commands and a manual review process to distinguish between malicious code and benign educational content. Every Friday, the catalog is automatically refreshed to include new skills and updated safety verdicts.
Integration is straightforward for any MCP-compatible client. Developers can clone the repository and add the server to their environment to search for and install skills on demand, such as requesting a specific skill for 3D camera movements.
What you can do with it
- 1Discover new agent capabilitiesSearch the entire catalog of 12,000 skills to find specific functional additions for your AI client.
- 2Automated skill installationAsk your AI agent to find and install a specific skill through a natural language command.
- 3Secure skill managementEnsure that any new skills added to your workflow have been vetted for prompt injection and malicious commands.
- 4Browse curated skill bundlesAccess high-quality, hand-picked skills through the top-picks marketplace.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableraveendi/skilloadout is public and reachablegithub.com/raveendi/skilloadout ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/raveendi/skilloadout ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubSkill Loadout 🎒
12,000+ AI agent skills — license-checked, safety-scanned, and refreshed every Friday.
Plug the whole catalog into Claude Code, Cursor, Codex CLI or Gemini CLI as a plugin marketplace or an MCP server. No website, no account, no backend: this repo is the product.
Get started in 30 seconds
Claude Code — add the curated top picks as plugins:
claude plugin marketplace add RaveenDi/Skilloadout
claude plugin install harness-engineering@skill-loadout
Any MCP client (Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI) — search and install from all 12,000+ skills on demand:
git clone --depth 1 https://github.com/RaveenDi/Skilloadout && cd Skilloadout/mcp-server && npm install
claude mcp add --scope user skill-loadout -- node "$(pwd)/index.mjs"
Then just ask your agent: "find me a skill for scroll-driven 3D camera moves and install it."
Why this one
- Safety first. Every skill is scanned for prompt injection, credential access, obfuscated
code, miners and destructive commands. Flagged skills get a reviewed verdict, not just a regex
match, so a YARA tutorial that mentions
xmrigisn't treated like malware. Skills reviewed as malicious are refused at install. See Safety. - License-respecting. Only redistributable licenses are mirrored, each with an
ATTRIBUTION.mdand the upstream license. Everything else is a link to its source. - Always fresh. Every Friday a GitHub Action re-syncs every source (updated skills refreshed, new ones added), and a Claude Code routine finds and reviews new skill repos and safety flags.
- Works everywhere. Skills follow the open Agent Skills format (
SKILL.md+ optional scripts/references) used by Claude (Code, Desktop, claude.ai), OpenAI Codex, Gemini CLI, GitHub Copilot and Cursor. Cursor rules are auto-converted toSKILL.md.
.
├── skills/ # the library — one folder per skill (mirrored verbatim + ATTRIBUTION.md)
├── originals/ # Skill Loadout Originals (authored here): 3D, fly-throughs, walkthroughs, context/harness/AI engineering…
├── catalog/
│ ├── sources.json # source registry (tiers, kinds); discovery agent appends to `discovered`
│ ├── index.json # generated catalog (one skill per line)
│ ├── top-picks.json # ⚡ curated picks — also what the plugin marketplace lists
│ ├── stacks.json # curated multi-skill bundles
│ ├── safety-verdicts.json, safety-overrides.json # reviewed safety verdicts (cached) + human overrides
│ ├── stats.json, changelog.json, discovery-log.json, state.json
├── scripts/
│ ├── sync.mjs # clone → find → license check → mirror/link → classify → dedupe → score → index
│ ├── discover.mjs # discovery agent (GitHub search + Claude review)
│ ├── verify-safety.mjs # Claude review of scan-flagged skills → benign / caution / malicious
│ ├── validate.mjs # catalog integrity checks
│ ├── build-marketplace.mjs # generates .claude-plugin/marketplace.json from top-picks.json
│ └── lib/ # taxonomy, license detection, safety scan, helpers
├── .claude-plugin/marketplace.json # Claude Code plugin marketplace (generated)
├── mcp-server/ # MCP server: search_skills, get_skill, install_skill, list_categories
└── .github/workflows/update-skills.yml # weekly (Friday) update agent
Quick start
npm install
npm run sync # download/refresh all sources (~10–20 min first time; incremental after)
Useful flags: node scripts/sync.mjs --only owner/repo, --force, --keep-cache.
npm run discover runs the discovery agent, npm run picks re-points the ⚡ Superpower list,
npm run validate checks catalog integrity, npm run marketplace rebuilds the plugin marketplace,
npm run verify-safety applies safety verdicts (and reviews new flags via the API if
ANTHROPIC_API_KEY is set).
To actually use the skills from an agent, see Connecting an agent directly below.
How the pieces work
Sync engine (scripts/sync.mjs)
- Expands
catalog/sources.json(repos + whole orgs such asgemini-cli-extensions). - For each repo:
git ls-remote→ skip if unchanged since the last run; else a blob-less, sparse, shallow clone that only downloads the skill folders (binary assets are skipped). - Finds
SKILL.mdskills, Cursor.cursorrules/.mdcrules and Geminigemini-extension.json. - License resolution per item: skill-folder LICENSE →
license:frontmatter → nearest parent LICENSE → repo metadata. Only redistributable licenses (MIT, Apache-2.0, BSD, ISC, CC0, CC-BY(-SA), MPL, GPL family…) are mirrored intoskills/<id>/with anATTRIBUTION.mdand the upstream license; everything else (proprietary, unknown, no license) is link-only — metadata + source link. - Classifies into 36 categories, runs a safety scan (prompt-injection text, credential access, obfuscated exec, destructive commands, pipe-to-shell), dedupes forks/mirrors by content hash and name+description, and scores quality (tier, stars, richness, freshness, safety).
- Writes
catalog/index.json,stats.json, and achangelog.jsonentry.
Weekly update (every Friday) — no API spend required
Friday morning — GitHub Action (.github/workflows/update-skills.yml, 00:23 UTC / 05:53 Sri
Lanka, or run it manually). Free on a public repo, no keys needed:
- Sync — incremental re-sync of every source: changed skills updated, new skills added, deleted ones removed.
- Safety verdicts — re-applies saved verdicts to the rebuilt catalog.
- Validate, rebuild the marketplace, commit and push.
Friday evening — /friday-update in Claude Code (skill),
on your Claude subscription. Claude Code does the review work in the session:
- Discovery —
discover.mjs --exportsearches GitHub for new skill repos and drops known ones and likely mirrors; Claude Code reviews each candidate's README and sample skills;--applyadds the accepted repos tosources.json. Re-review — sources an old popularity heuristic auto-accepted are vetted 25 a week, biggest first (--export-rereview/--apply-rereview); rejected ones getstatus: "rejected". - Sync — mirrors the newly accepted sources (and anything changed since the morning) and drops
rejected sources' skills;
validate.mjscounts those removals as deliberate, not a failed sync. - Safety review —
verify-safety.mjs --evidence-onlyexports flagged skills that still need a verdict; Claude Code reviews them in batches;--applystores each batch. - Validate, rebuild the marketplace, commit and push.
Optional: add an ANTHROPIC_API_KEY repository secret and the Action does both reviews itself via
the API (roughly $12–25/month). Without it, the Action skips discovery and leaves new flags
unverified until the Friday-evening run.
Categories & Superpowers
catalog/taxonomy.json defines 36 categories in 5 groups (Build, Create, Business, Work, Knowledge) —
the single source of truth for the sync classifier. catalog/top-picks.json holds the hand-curated
⚡ Superpower skills per category (the best few) — this is also exactly what the plugin
marketplace lists. Bulk/auto-generated collections are listed in sources.json → exclude so only
top skills make it in.
Connecting an agent directly (no website)
Two ways to pull skills straight from this repo, both regenerated by the scheduler on every sync — no server to run or host:
-
Claude Code plugin marketplace —
.claude-plugin/marketplace.jsonlists the ⚡ curated top picks (catalog/top-picks.json) as installable plugins; each is just its existingskills/<id>/folder (a rootSKILL.mdalready loads as one skill, no wrapping needed):claude plugin marketplace add RaveenDi/Skilloadout claude plugin install <skill-id>@skill-loadoutRegenerate manually with
node scripts/build-marketplace.mjs(the scheduler does this every run). -
MCP server (
mcp-server/) —search_skills/get_skill/install_skill/list_categoriestools over the entire catalog (not just top picks), for Claude Code, Claude Desktop, Cursor, Codex CLI, or Gemini CLI. Readscatalog/index.jsonandskills/directly off disk — no network, no key. This repo's own.mcp.jsonwires it in for anyone who opens the repo in Claude Code; seemcp-server/README.mdfor other clients.
Safety
Skills are instructions plus scripts that an agent will run, so a bad one is a real supply-chain risk. Two passes protect you:
- Scan (
scripts/lib/safety.mjs) — on every sync, each skill's files are checked for prompt-injection text, credential-file access, obfuscated execution, crypto-miners, destructive commands, pipe-to-shell installs and disabled safety checks. - Review (
scripts/verify-safety.mjs) — a regex can't tell a skill that detects an attack from one that performs it, so each flagged skill's matched lines go to Claude (in the/friday-updateClaude Code routine, or via the API if a key is set) for a verdict: benign, caution (legitimate but read it before running) or malicious. Verdicts are cached incatalog/safety-verdicts.jsonby a hash of the matched content, so a skill is re-reviewed only when that content changes.
The MCP server shows only reviewed warnings. Unreviewed matches are labelled unverified, never as
findings. install_skill refuses anything reviewed as malicious, and the plugin marketplace leaves
it out.
Maintainers: if your skill got a wrong verdict, open an issue. Human verdicts live in
catalog/safety-overrides.json and take priority over the automated review while the matched
content stays unchanged.
Installing downloaded skills
| Tool | Location |
|---|---|
| Claude Code | ~/.claude/skills/<name>/ or .claude/skills/<name>/ in a repo |
| Claude.ai / Desktop | Settings → Capabilities → Skills → upload the zip |
| Codex / ChatGPT | ~/.codex/skills/<name>/ |
| Gemini CLI | ~/.gemini/skills/<name>/ or .gemini/skills/<name>/ |
| GitHub Copilot | .github/skills/<name>/ |
| Cursor | .cursor/skills/<name>/ (rules → .cursor/rules/) |
Adding sources
Edit catalog/sources.json → sources ({ "repo": "owner/name", "tier": 1|2|3, "origin": "…" },
optional include path prefixes, exclude regexes, kind: "rules" | "gemini-extension") and run
npm run sync -- --only owner/name. Put known mirrors/aggregators in exclude.
Licensing
Skill Loadout's own code and originals/ are MIT. Mirrored skills keep their upstream licenses — see
each folder's ATTRIBUTION.md and LICENSE. Link-only skills are not redistributed.
Skilloadout MCP questions
How do I install skills using an MCP client?
You can search and install skills on demand by cloning the repo and adding the MCP server to your client like Claude Desktop or Cursor.