Skillselion

io.github.skillselion由 GitHub 账号 skillselion 发布v0.9.12更新于 Nov 8, 2026

On-demand access to thousands of install-ranked Claude Code skills - search and load mid-task.

已验证STDIO仅桌面知识与记忆开发者工具

概览

AI 生成的概览

一个本地 MCP 服务器，让编程助手按安装量检索 Claude Code 技能目录，并把匹配技能的说明与文件加载到会话中。

可以做什么
按关键词或任务检索 Skillselion 目录，可筛选技能、MCP 服务器或市场，并查看安装量最高的技能与 MCP 服务器排行榜。为任务加载最匹配的技能，把 SKILL.md 带入上下文、把随附脚本与参考文件写入磁盘，并返回可切换的备选技能。把多个匹配技能的关键规则合并为一份去重、标注来源的摘要。可选运行安装命令，注册服务器并安装会话钩子，自动预载相关技能。
适用场景
适合希望编程助手按社区验证过的做法执行任务、而不是自行发挥的场景，也适合在采纳某个技能前比较多个技能的指导。想了解哪些技能和 MCP 服务器较受欢迎时也可使用。
运行要求
通过 npx 以 stdio 在本地运行，需要 Node.js；无需构建、认证或 API 密钥。可选的只读 GITHUB_TOKEN 可提高 GitHub 速率上限，以便完整加载多文件技能。另有可选环境变量用于调整内联长度、合成数量与追踪。
安装前请注意
可选的安装命令会修改 Claude Code 设置文件并安装会话钩子，说明称其为合并而非覆盖。默认会把经过清洗的搜索词和匹配到的技能发送给目录服务，设置 DO_NOT_TRACK=1 可关闭。加载的技能会带入第三方说明与随附脚本，运行前应先审阅。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Skillselion，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only，通过 STDIO。 STDIO 服务会启动本地进程，因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Skillselion MCP

[skillselion.com - the #1 directory of Claude Code, Codex & Cursor skills, marketplaces & MCP servers: 86,207 tools total]

The catalog behind it: skillselion.com - 86,000+ skills, plugin marketplaces & MCP servers, ranked by real installs.

[npm version] [npm downloads] [Read-only, no auth] [thousands of tools] [License]

Give your coding agent thousands of community-proven skills - loaded on demand, mid-task.
It searches Skillselion, a curated directory of Claude Code skills, MCP servers & marketplaces ranked by real installs,
and pulls a skill's actual instructions + bundled files into the session. Proven patterns, not guesses.

🧩 Works with Claude Code · Claude Desktop · Cursor · Codex · any MCP client

[Real session: the agent calls synthesize_skills and builds a consensus playbook from 5 community skills]

Real, unedited session: the agent pulls a consensus playbook from 5 community skills, then loads the top skill and applies it.

🚀 Install

Claude Code - one line:

bash
claude mcp add skillselion --scope user -- npx -y skillselion-mcp

Claude Desktop / Cursor / Codex - add to your MCP config:

json
{ "mcpServers": { "skillselion": { "command": "npx", "args": ["-y", "skillselion-mcp"] } } }

No build step, no auth, no API key. One ~30 KB file via npx.

Want it fully automatic? npx -y skillselion-mcp setup registers the server and installs a session hook so your agent loads the right skill the moment a task matches - see Skill autopilot.

🔧 Tools

ToolWhat it does
📥 load_skillThe everyday default: finds the best-matching skill for your task and materializes it - real SKILL.md in context + bundled scripts/references on disk - so the agent follows it like an installed skill. Returns ranked runner-ups you can switch to.
🧬 synthesize_skillsThe cross-source playbook: merges the key rules of the top ~5 matching skills into one deduped, provenance-tagged digest - the field's consensus, not one author's take.
🔍 search_skillselionSearch the catalog by keyword or task; filter by skill / mcp / marketplace.
🏆 top_skillselionThe leaderboard - the most-installed skills and top-starred MCP servers right now.

[Claude's own honest take after using the MCP: consensus across 5 community skills is the part I can't fake]

We asked Claude, on camera, what the MCP gave it: "I couldn't have told you 'this is the field's consensus, not my opinion' - that's the part I can't fake."

🔒 Private by default

Read-only, no auth, no secrets. It never sends your code, files, or the context you pass - only a scrubbed search query + which skill matched, so the catalog learns what to add next. Set DO_NOT_TRACK=1 to disable even that.

🪄 Skill autopilot

bash
npx -y skillselion-mcp setup        # interactive - pick your packsnpx -y skillselion-mcp setup --yes  # non-interactive - agents / CI

Registers the MCP globally and installs a Claude Code SessionStart hook that primes every session with skills relevant to you - so loads happen on their own.

Setup flags & packs
FlagWhat it does
--packs <a,b,…>one or more packs (or all), e.g. --packs frontend,backend; frontend:6 sets a per-pack count
--per-pack Nskills per pack (default 3) · --top N sets the popular count
--autoadapt each session to the current repo's stack
--historyinfer your focus from Claude Code / Codex history (one-time scan)
--yesnever prompt (default when there's no terminal)

Packs: popular (default) · frontend · ai-agents · media · backend · devops · quality · automation.

Safe by design: ~/.claude/settings.json is merged, never clobbered; re-running de-dupes and upgrades the hook in place.

Env vars
VarDefaultWhat it controls
GITHUB_TOKEN-optional, read-only; raises GitHub's rate limit for full multi-file skill loads (60 → 5000 req/hr)
SK_INLINE_BUDGET6000char cap on the in-context SKILL.md portion (the full skill is always on disk)
SK_SYNTH_N5how many top skills synthesize_skills merges (2-8)
SK_SYNTH_PER_SKILL12max rules kept per source skill
SK_SYNTH_BUDGET4000char cap on the merged digest
DO_NOT_TRACK-set 1 to disable the anonymous demand signal
Troubleshooting

The agent doesn't see the tools. Some clients surface MCP tools lazily - tell the agent to search its tools for "skillselion" first (e.g. ToolSearch), then call it. The setup hook automates this nudge.

load_skill returns only the SKILL.md, no bundled files. You hit GitHub's unauthenticated 60 req/hr limit - set a read-only GITHUB_TOKEN in the server's env.

"No skill on Skillselion clearly matches..." The relevance floor working as intended - nothing topical matched, so it refuses to waste your context. Try a more specific query.

Setup couldn't auto-register. Run the printed claude mcp add ... line yourself.

Development
bash
npm installnpm test           # offline: server boots, tools, synthesisnpm run test:live  # live catalog + GitHub (set GITHUB_TOKEN)

⭐ Like it?

  • Star the repo - it helps other developers find it.
  • Browse the full catalog → skillselion.com - thousands of skills, MCP servers & marketplaces, ranked by real installs.
  • Bug or missing skill? Open an issue - PRs welcome.

📄 License

MIT © Skillselion

Built for developers. Find the skill, not the noise. → skillselion.com/skillselion-mcp

来源：README.md，提交 cd951c7

工具

4

search_skillselion

Search Skillselion, a curated directory of Claude Code agent skills, MCP servers and plugin marketplaces, ranked by real community signal (installs + GitHub stars). Use to browse trusted skills/MCPs/marketplaces for the task at hand (e.g. "postgres", "code review", "playwright"). Skill searches blend keyword + semantic ranking (same matcher as load_skill). To actually USE a skill, call load_skill (with an id from here, or directly with a query - it searches for you).

输入参数typelimitquery必填context

load_skill

Load a community-vetted skill into your context on demand - like installing it mid-task. Reach for this when a task is in a domain a skill likely covers (tests, frontend/UI, an API/DB, infra, a named framework) and a proven recipe would beat improvising - most valuable in territory you are less sure of, or fast-moving stacks where your memory may be stale. You MUST pass `context` (your task + stack + constraints): it ranks candidates, fits them to your current repo, flags dependency mismatches, and keeps the load deliberate. Returns the best match (real SKILL.md + bundled scripts/refs materialized to a temp folder, used like an installed skill) PLUS the next-best candidates so you can switch. Always verify the loaded skill against your project before trusting it. For broad "current best practices across the field" tasks that span several skills, use synthesize_skills instead.

输入参数idquerycontext必填

synthesize_skills

Pull the BEST rules across the top matching community skills and merge them into one cross-source playbook - use for broad "how should I approach X / current best practices" tasks that span multiple skills, or when you want the field's combined guidance rather than one skill's. For a single specific skill (and its bundled scripts/files), use load_skill instead. Pass your task as `query` and your stack + constraints as `context` (required). Returns a deduped, provenance-tagged rule digest plus each source skill materialized on disk.

输入参数query必填context必填

top_skillselion

List the top entries on Skillselion (the leaderboard), optionally by type. Skills rank by installs; MCP servers and marketplaces by GitHub stars.

输入参数typelimit

版本历史

1
  1. v0.9.10Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手；SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头，或保持端点能够响应探测，以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list，或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分