Where to find Claude skills and MCP servers

A quick word on our interest: we run Skillspace, one of the directories below, and we've tried to describe it the way we describe the others, weaknesses included. Every count on this page was read from the directory's own site on October 31, 2026, except the official MCP Registry figure, which we counted ourselves through its public API that day.

The short list

Skills directories

Claude's built-in directory

Inside Claude, Customize then Skills lists what you have, grouped by created by you, from your organisation, shared with you, and from Anthropic and partners, with a Discover tab for more (Anthropic docs). Partner skills are built to work with those partners' connectors, and on Team and Enterprise plans owners can publish organisation skills with publishing set to open or requiring review. It's the safest place to start and the smallest; its job is to cover common work well, not to be exhaustive.

Anthropic's skills repository

The anthropics/skills repository holds Anthropic's example skills (creative and design, development, enterprise and communication) and the document skills for Word, PDF, PowerPoint and Excel, which are source-available rather than open source. Most of the rest is Apache 2.0. In Claude Code, /plugin marketplace add anthropics/skills installs it as a marketplace. Anthropic labels these "for demonstration and educational purposes", which makes the repo the best place to learn how a good skill is structured, and a template to copy.

skills.sh

Vercel's skills.sh is a leaderboard of skills ranked by installs, with all-time, trending and hot views, an install command (npx skills add owner/repo) and a list of 20 supported agents including Claude Code, Cursor, Codex, GitHub Copilot and Gemini. Install counts are a strong signal of what's useful and a weak signal of quality, since popular skills can still be badly written, but it's the fastest way to see what other people rely on. It has a security audits section, whose method we didn't review for this page.

Skillspace (ours)

Skillspace indexes public agent skills from GitHub (139,726 when we checked), merges duplicates by keeping the copy from the most-starred repository (4,310 merged), sorts them into 276 categories and grades each from A to F with a published scoring method. The grade weighs how a skill is built (whether its description will trigger it correctly, instruction quality, modularity, progressive disclosure) and maturity signals such as references, scripts, a licence, versions and evals. A static safety scan can cap the grade, and skills a model flags for prompt injection or similar are excluded from rankings. 27,503 skills pass as launch-quality: graded A to C with a clean scan. A free MCP server at https://mcp.skillspace.sh/mcp lets an agent search those by the problem it needs solved, with no sign-in, limited to 20 requests a minute and 1,000 a day per client.

Its weaknesses, several of which its own methodology page states:

• The grade measures construction, not results. The methodology says plainly that it "cannot tell you the skill works". Only 56 skills have blind head-to-head tests on real tasks, and 1,812 carry a judge-model review.
• The scan is static. It reads files with rules and never runs anything, so "a clean security verdict means nothing matched, not that nothing is there", and its rules produce false positives (one fired on 133 of 396 benchmark skills, most harmless).
• Ratings are snapshots. The current one was generated on September 2, 2026, so a skill changed since then is graded on its old version.
• No public submission and no install button on the site. Skills come from GitHub indexing only, and the team product (grading and distributing a company's own skills) is in closed beta.

Use it when you're searching a crowded category and want a first filter, then read the skill before installing, the same as anywhere else. If you only need the most common skills, Anthropic's directory or skills.sh will get you there faster.

MCP server registries

The official MCP Registry

The official MCP Registry is maintained in the open by MCP contributors and is the canonical place a server's publisher lists its name, version, npm or PyPI package and remote endpoint. We paged through its public API on October 31, 2026: 196,993 entries across all versions, 52,715 servers whose latest version is active, 35,177 of them with a remote (hosted) endpoint and 21,134 with an installable package. It has no ratings, reviews or usage data, so it answers "what's the real, current endpoint for this server" well and "which one is good" not at all. Other directories and clients use it as a source.

Glama

Glama's directory is the largest we found, at 112,648 open-source servers, with letter grades for licence, quality and maintenance, sorting by recent usage, downloads and GitHub stars, and filters for language, hosting (44,176 remote, 42,303 local, 21,295 hybrid) and capability. The grades make it the best place to compare several servers for the same system. Its size also means a lot of abandoned and duplicate entries, so lean on the maintenance grade.

Smithery

Smithery lists 27,365+ servers and focuses on using them: CLI commands such as npx smithery mcp add notion, publishing for server authors, and OAuth flows and credential injection handled for you. One thing to know this week: its site banner says its MCP Proxy and managed identity shut down on October 31, 2026, so if you relied on those, check what replaces them before you plan around Smithery-hosted connections.

mcp.so

mcp.so describes itself as a community-driven platform for third-party MCP servers, with sections for servers, clients, CLI tools and skills, a featured list it calls hand-picked and production-ready, rankings, and open submission with review before listing. It doesn't name its operator or publish a total, and it has no quality ratings beyond featured picks and star counts. It's a good browse for discovery; confirm the details against the server's repository or the official registry.

How to choose one safely

Whatever directory you use, the safest connector is usually the one the vendor of the system publishes: a CRM's own MCP server, your database vendor's, your ticket tracker's. Check that first, because it'll track the vendor's API and permissions model. Then apply the same rules everywhere:

1. Read before you install. A skill is instructions and possibly scripts that run in your session; an MCP server is code with your credentials. Anthropic's docs tell you to read a skill's files before turning it on.
2. Prefer read-only, narrow scopes and service accounts for any connector, and add write access only when a task needs it.
3. Pin versions for anything a team depends on, so an upstream change doesn't silently alter behaviour.
4. Prefer maintained over popular. A recent commit and an issue tracker with replies matter more than stars.

If you aren't sure you need a skill or a server at all, the decision tool will tell you which rung your task sits on, and the explainer covers what each one is.