How to install Claude skills and plugins in any client

The fastest way to a working skill with real data

superagnt is a marketing plugin for Claude and ChatGPT that helps businesses find customers through outbound, SEO and ads. Its plugin is the example in every section of this page: one install gives your AI 8 skills, and one sign-in gives them live data. A skill on its own only has instructions; this is the shortest install in each client that ends with both.

claude plugin marketplace add superagnt/plugins
claude plugin install superagnt@superagnt

claude mcp add --scope user --transport http superagnt https://mcp.superagnt.com/mcp
claude mcp login superagnt

codex plugin marketplace add superagnt/plugins
codex plugin add superagnt@superagnt
codex mcp login superagnt

openclaw plugins install --marketplace superagnt/plugins superagnt --force --accept-capabilities
openclaw mcp add superagnt --url https://mcp.superagnt.com/mcp --transport streamable-http --auth oauth
openclaw mcp login superagnt

{
  "mcpServers": {
    "superagnt": {
      "url": "https://mcp.superagnt.com/mcp"
    }
  }
}

Every client side by side

Seven clients, seven questions. A tested cell is our install from an empty profile on Nov 5, 2026, with the client's version and the seconds from the first command to the first tool result. The rest say what the client's own docs say, with the day we read them, and that we haven't run them.

Install superagnt's plugin in each client

Each tested step shows what our run printed and how long it took. Your commands point at superagnt's public server; ours pointed at our test server, which is the only difference you'll see in the output, except in Codex, where we signed in with codex mcp add to do that.

| keyword | search_volume | cpc | competition | difficulty | intent | volume_trend_yearly |
| claude skills | 40500 | 4.81 | LOW | 27 | informational | 45150 |

| keyword | search_volume | cpc | competition | difficulty | intent | volume_trend_yearly |
| claude skills | 40500 | 4.81 | LOW | 27 | informational | 45150 |

| keyword | search_volume | cpc | competition | difficulty | intent | volume_trend_yearly |
| claude skills | 40500 | 4.81 | LOW | 27 | informational | 45150 |

Skills, plugins and connectors

In Anthropic's words, connectors give Claude access to a tool or data source, skills teach it how to do a task the way you do it, and plugins package skills, connectors, commands and agents into one unit you install once.

• A skill without data is advice, and an MCP server without a skill is a box of tools. A marketing skill that looks up keywords, ads or people needs both, which is why superagnt's plugin ships its skills and its server together.
• Claude picks a skill by its description, so many installed skills cost little: Claude Code gives the whole skill list about 1% of the context window and, when it overflows, drops the descriptions of the skills you use least.
• To use a skill on purpose, call it: type /skill-name in Claude Code (/plugin-name:skill-name for a plugin's), pick it from the / menu in the Claude app's message box, type $skill-name in Codex, or ask for it by name ("use the super-seo skill to…").
• In Claude Code, custom slash commands are now skills: .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both make /deploy. Anthropic's split: CLAUDE.md is context loaded every conversation, a skill loads when needed, a subagent is a worker with its own context, and a hook runs every time its event happens.
• One format, many clients: a skill is an Agent Skills folder, which Claude, Claude Code, Codex, ChatGPT, Cursor and OpenClaw all read, each from its own place (the matrix's last columns). A skill installed for one client shows up in another only where they share a folder: Cursor reads ~/.claude/skills, while Codex and OpenClaw import from Claude Code rather than reading it.
• Your Claude account follows you: skills, plugins and connectors added on claude.ai or in the desktop app reach the web, Desktop, mobile and Cowork, and sync into Claude Code 2.1.273 or later when it's signed in with the same account. Nothing goes the other way: what you add from the Claude Code command line stays on that machine.
• Anthropic's own skills for Excel, Word, PowerPoint and PDF come built in: with code execution on, Claude uses them by itself, and each skill has an on and off switch in Customize > Skills. Skills you turn on also work in the Claude add-ins for Excel, PowerPoint, Word and Outlook (paid plans), where you type / to pick one.
• What a skill's code can reach online is the Allow network egress setting, under Settings > Capabilities; Team plans start at package managers only, and new Enterprise organizations with it off. Connectors aren't limited by it, which is how a skill gets live data with egress off.
• Developers can send skills with API calls too: upload a custom one with POST /v1/skills and list up to 20 per request in container.skills, run by the code execution tool. The Agent SDK reads skills from .claude/skills folders instead.

Make your own skill

A skill is a folder named for the skill, with a file called exactly SKILL.md inside. The file opens with two fields, then the instructions:

---
name: keyword-brief
description: Writes a one-page brief for a target keyword (search volume, intent, what ranks). Use when someone asks for a keyword brief or what to write for a keyword.
---

# Keyword brief

1. Look up the keyword's monthly volume and intent.
2. Read the top results and note what they cover.
3. Write the brief: the angle, the sections, the questions to answer.

keyword-brief/
  SKILL.md        the name, the description, the instructions
  scripts/        optional: code the skill runs
  references/     optional: docs it reads when it needs them
  assets/         optional: templates and files it uses

• The name is lowercase letters, numbers and single hyphens, up to 64 characters, and matches the folder. Anthropic's pages disagree on the description's limit (200 characters in the help center, 1,024 in the docs and the spec), so put the useful part in the first 200. Keep SKILL.md under 500 lines and move the rest to references/.
• The description should say what the skill does and when to use it, in the words people use when they ask for it. One that never fires is usually vague; one that fires too often, too broad.
• Outside Claude Code, only the spec's six fields work: name, description, license, compatibility, metadata, allowed-tools. Anything else fails the upload with Unexpected key(s) in SKILL.md frontmatter. Claude Code adds its own, such as when_to_use, disable-model-invocation and context: fork, and ignores fields it doesn't know.
• By hand in Claude Code: mkdir -p ~/.claude/skills/keyword-brief, write the file, and the running session picks it up; /skills lists what loaded and /reload-skills re-scans the folders. In the Claude app, zip the folder (the folder itself at the top of the ZIP) and upload it under Customize > Skills; on Pro, Max, Ultra or Team, Cowork on a Mac can also record your screen and let Claude write the skill.
• Or let Claude write it: Anthropic's skill-creator skill drafts the folder from a conversation. On claude.ai, turn it on under Customize > Skills and ask Claude to use it; in Claude Code, /plugin install skill-creator@claude-plugins-official. Anthropic's guide expects about 15 to 30 minutes to build and test a first working skill.
• Test it in fresh sessions: a few realistic asks with the skill on, the same asks with it off, and compare. Then ask for unrelated things to check it stays quiet, tighten the description and repeat. skill-creator can generate those should-fire and shouldn't-fire prompts and measure the hit rate.

A real one: superagnt's Super SEO skill

superagnt's super-seo skill has the same shape: a SKILL.md whose description lists the asks it answers ("run my SEO", "improve my rankings", "check my AI visibility"), and a references/ folder it reads when it needs it. It installs on its own with npx skills add superagnt/plugins --skill super-seo, or with its plugin from the same marketplace, and it does its research with superagnt's live search data.

Share it as a plugin

• A plugin is a folder with its manifest at .claude-plugin/plugin.json and its skills under skills/. Try it for one session with claude --plugin-dir ./my-plugin, and check it with claude plugin validate ./my-plugin.
• A marketplace is a repo whose .claude-plugin/marketplace.json names it, its owner and its plugins. Push it to GitHub and anyone can add it: /plugin marketplace add owner/repo in Claude Code, Add marketplace in the Claude app, codex plugin marketplace add owner/repo in Codex, openclaw plugins install --marketplace owner/repo <plugin> in OpenClaw. superagnt's repo, superagnt/plugins, carries Claude's and Codex's marketplace files side by side, which is how one repo installs in all three.
• To list it in Anthropic's directory, submit the GitHub repo from a paid Claude plan at claude.ai/directory/manage; Anthropic checks each submission first.
• For a team: commit .claude/skills to the repo, or run claude plugin marketplace add your-org/your-marketplace --scope project once and commit the .claude/settings.json it writes. On Team and Enterprise, owners add skills and plugins for everyone under Organization settings > Plugins & skills, and can make a plugin available, installed for everyone, or required.

Where to find skills, and how to check one

• Anthropic's: the directory in the Claude app (Customize > Skills > + > Browse skills, and the Discover tab under Plugins), the anthropics/skills repo (/plugin marketplace add anthropics/skills), and the official plugin marketplace, which Claude Code adds by itself in your first interactive session and lists under /plugin's Discover tab.
• OpenAI's: $skill-installer in Codex, and the plugin directory under Plugins in ChatGPT.
• OpenClaw's: ClawHub (openclaw skills search <words>), which showed us a security audit for the skill before installing it.
• Any client: npx skills add owner/repo installs from any GitHub repo of skills (Node 22.20 or later).
• Marketing skills that come with their data: superagnt's marketplace, superagnt/plugins, whose skills run on live search, ads, social and people data from its server.

Check a skill before you install it

• Anthropic's rule for the Claude app: only install skills from trusted sources, and review one from anywhere else, a colleague included, before turning it on.
• Read what runs: SKILL.md and its scripts, and for a Claude Code plugin also hooks/hooks.json, .mcp.json and bin/. A plugin can run any code with your user's permissions, whichever marketplace it comes from.
• Watch skills that fetch web pages: fetched content can carry instructions (prompt injection). The skills CLI says it too as it finishes: "Review skills before use; they run with full agent permissions." Enterprise plans can have Anthropic scan third-party skills and plugins before they run.
• Many lists of the best skills come from directories and vendors, so check the repo itself: when it last changed, what it runs, and what data or accounts it needs.
• One instruction some guides give is wrong: copying a SKILL.md into ~/.claude/plugins/<name>/. That folder is where Claude Code keeps installed plugins, not a skills folder; a personal skill goes in ~/.claude/skills/<name>/.

What broke, and the fix

Every failure our runs and checks hit, each dated, with the message word for word (trimmed where it says …) and what fixed it.

1. Claude CodeOur check, Nov 4, 2026, Claude Code 2.1.318

   You asked the agent to sign in to an MCP server for you, and the sign-in failed.

   Couldn't complete authentication for "superagnt": stdin isn't a terminal, so authentication can't be completed here. Re-run in an interactive terminal…

   fix Run claude mcp login superagnt yourself, in a terminal, and approve the consent screen it opens. Inside a Claude Code session, /mcp does the same.

2. Claude CodeAnthropic's docs, read Nov 4, 2026; Our run, Nov 5, 2026, Claude Code 2.1.318

   A connector you added on claude.ai doesn't show up in Claude Code.

   fix Run claude auth status. If authMethod says api_key, claude.ai connectors never load (Anthropic's Claude Code docs), so add the server directly: claude mcp add --scope user --transport http superagnt https://mcp.superagnt.com/mcp, then claude mcp login superagnt. All three of our Claude Code runs signed in this way.

3. Any client (the skills CLI)Our check, Nov 4, 2026, skills 1.7.6 on Node 20.18.1

   npx skills add superagnt/plugins --skill super-seo stops with a stack trace.

   npm warn EBADENGINE Unsupported engine { … required: { node: '>=22.20.0' }

   fix Install Node 22.20 or later, then run it again.

4. CodexOur check, Nov 4, 2026, Codex 0.166.0 (also Oct 17 on 0.159.3 and 0.163.0)

   You added one of superagnt's skill packs and its skills load, but no superagnt tools show up.

   No MCP servers configured yet.

   fix Codex doesn't install a plugin's dependencies, so add the base plugin first, as the lines for a pack do: codex plugin marketplace add superagnt/plugins, then codex plugin add superagnt@superagnt, then codex plugin add super-seo@superagnt.

5. CodexOur run, Nov 5, 2026, Codex 0.166.0

   codex mcp list shows the plugin's superagnt server as not logged in.

   superagnt  …  enabled  Not logged in

   fix Sign in once: codex mcp login superagnt. It opens superagnt's consent screen in your browser.

6. CodexOur runs, Nov 4, 2026, Codex 0.166.0

   codex exec answers without superagnt's tools, though codex mcp list shows the server signed in.

   fix In codex exec, the first turn reached the model a fraction of a second before the server finished connecting, so that turn had no superagnt tools, and a resumed exec did the same. Start codex in a terminal and ask once it's ready: every interactive run had the tools from the first turn.

7. OpenClawOur install, Nov 4, 2026, OpenClaw 2026.10.10 on Node 20.18.1

   Installing or updating OpenClaw fails before it starts.

   [openclaw] error: this OpenClaw release requires Node >=24.16.0 <25 || >=26.1.0.

   fix Install Node 24.16 or later on the 24 line, or 26.1 or later (Node 25 isn't supported), then install OpenClaw again.

8. OpenClawOur check, Nov 4, 2026, OpenClaw 2026.10.10

   Installing a plugin from a Claude marketplace stops before it installs anything.

   Install cancelled; rerun with --force after reviewing the source.

   fix The source isn't in ClawHub's review, so OpenClaw asks you to confirm it: add --force once you trust the publisher.

9. OpenClawOur check, Nov 4, 2026, OpenClaw 2026.10.10

   With --force, the install stops a second time.

   Plugin "superagnt" requires capability consent. The plugin was not installed.

   fix Add --accept-capabilities too. The full line: openclaw plugins install --marketplace superagnt/plugins superagnt --force --accept-capabilities.

10. OpenClawOur run, Nov 5, 2026, OpenClaw 2026.10.10

    openclaw skills install @superagnt/seo-data installs the skill, then warns that the Gateway didn't refresh.

    Skill files changed, but local Gateway refresh failed: … Restart the Gateway to refresh skills.

    fix The skill is installed. If you run OpenClaw's Gateway, restart it to load the skill (openclaw gateway restart for the installed service). Without a Gateway there's nothing to do: our run's next openclaw agent --local session had the skill.

11. OpenClawOur run, Nov 5, 2026, OpenClaw 2026.10.10

    A ClawHub skill installed fine, but the agent never sees it.

    Visible to model: no

    fix openclaw skills info <skill> says what the skill is missing. superagnt's ClawHub skills need your API key, so save it: openclaw config set skills.entries.seo-data.apiKey <YOUR_API_KEY>.

12. OpenClawOur run, Nov 5, 2026, OpenClaw 2026.10.10, once

    After Authorize, the browser never came back and the sign-in kept waiting.

    Waiting for the browser to return to OpenClaw...

    fix Cancel it and run openclaw mcp login superagnt again; our second try signed in. If your browser can't reach OpenClaw at all (a remote machine), OpenClaw prints its own way out: openclaw mcp login superagnt --code <code>, with the code from the address your browser lands on.

13. Claude DesktopOur check, Oct 5, 2026 (not rerun since)

    A connector you added in the browser isn't in the Claude desktop app you already had open.

    fix The app loads your account's connectors when it starts: refresh it (Cmd+R) or restart it.

14. Claude DesktopOur check, Oct 5, 2026 (not rerun since)

    A claude:// link to add a connector opens the app, but the dialog isn't filled in.

    fix Use the https link (claude.ai/customize/...): it keeps the name and URL and opens in your browser.

What the vendors' docs list too

Failures our runs didn't hit but each client's own docs describe, with their fix.

How we tested

• Nine installs on Nov 5, 2026: Claude Code 2.1.318, the Codex CLI 0.166.0 and OpenClaw 2026.10.10, each installing a skill on its own, superagnt's plugin, and the MCP server alone. The clients were the current releases from npm that day.
• An empty profile each time: a new home folder and an empty config folder for the client, on a Linux machine. Nothing was set up ahead except the client's model sign-in.
• The clock ran from the first install command to the moment the client handed the first tool result back. It includes the sign-in, where a browser already signed in to superagnt approved the consent screen, and it leaves out your own reading and typing.
• A script played the model's side of each session, so no model's thinking time is in the numbers; the client, the install, the sign-in and the tool call were all real. Codex ran in its interactive terminal, the way you'd use it, with the question typed once it was ready.
• The MCP server was superagnt's test server, with a new account, which is why the commands we ran show <server url> where yours have superagnt's public URL. The first data call was a keyword lookup and cost 2.2 cents of the account's free trial.
• We couldn't run the Claude app, Cowork, ChatGPT or Cursor here: each needs a signed-in account in a browser or a desktop app. Their cells say so and quote each vendor's docs, read on the day each cell gives.

Common questions