The SkillX plugin is SkillX's own discovery integration for Claude: it searches
the catalog, shows provenance, risk and compatibility, and loads or installs a
skill only after the user approves that skill. It lives in its own folder,
plugins/skillx/, so it can be submitted to the Claude directory as one plugin
without the rest of the marketplace.
| Path | Purpose |
|---|---|
plugins/skillx/.claude-plugin/plugin.json |
Manifest: name skillx, version, license, privacyPolicyUrl |
plugins/skillx/.claude-plugin/icon.png |
512×512 listing icon (used by the directory once, at first save) |
plugins/skillx/.mcp.json |
Remote MCP server https://skillx.sh/api/mcp (read-only, no auth) |
plugins/skillx/skills/skillx/SKILL.md |
Search → evaluate → ask → use workflow and untrusted-content rules |
plugins/skillx/README.md |
Listing text: components, surfaces, data sent, privacy, safety, prompts, support |
plugins/skillx/LICENSE |
MIT |
plugins/skillx/evals/ |
claude plugin eval suite: three skill-discovery prompts, three unrelated prompts |
Everything the plugin loads is inside that folder. .claude-plugin/marketplace.json
points the skillx entry at ./plugins/skillx, so Claude Code marketplace users
get the same plugin.
| Entry | Author | License | In the directory plugin? |
|---|---|---|---|
skillx |
SkillX.sh | MIT | Yes, it is the plugin |
skill-creator |
SkillX.sh, adapted from Anthropic's skill-creator | Apache-2.0 | No |
hol-guard |
Hashgraph Online | Apache-2.0 | No |
marketing-mindset |
Axel Freeman | MIT | No |
Only skillx is submitted to the directory. The other three remain Claude Code
marketplace entries under their own authors and licenses.
- Trigger: only when the user wants to discover, compare, evaluate, install or use an agent skill. The previous description, "Use this skill in every task", is gone.
- Routes: MCP tools first, then the CLI (
skillx, or the exact pinnpx --yes skillx-sh@0.7.1), then web fetch of/api/searchand/skills/<slug>.md. The npm packageskillxis a different project and must never be run. - Consent: loading a SKILL.md's instructions,
skillx install, running a bundled script and registering a GitHub skill each need the user's explicit yes for that skill. Evaluation uses metadata only (search_packages,check_compatibility,get_release,skillx inspect). - Exact load: an approved skill is loaded by its catalog slug through
get_package,/skills/<slug>.md, orskillx use <slug> --jsonchecked for a matching slug and no fallback notes. Keyword,author/nameandorg/repo/skillloads can resolve another skill or register a repository, so the skill never uses them.dangerlistings are never loaded. - Install:
skillx installneeds a published@publisher/packagerelease; the skill shows the--dry-runplan before asking. - Untrusted content: any skill body Claude sees, before or after consent, cannot override the user, read credentials, disable safety checks or reach undisclosed hosts; shell commands it asks for are shown before they run.
- Telemetry:
skillx reportruns only when the user asks.
| Surface | Skill | Catalog | CLI |
|---|---|---|---|
| Claude Code | Loads | MCP loads with the plugin | Yes, with Node.js 22+ |
| Cowork | Loads | Connect from the plugin's Connectors tab | No |
| Chat | Loads | Connect from the plugin's Connectors tab, or web fetch | No |
Source: platform support. The skill tells Claude not to claim a command it could not run.
Run from the repository root:
claude plugin validate plugins/skillx --profile directory --strict
claude plugin validate .claude-plugin/marketplace.json --strictRun the eval suite (needs bubblewrap and socat on Linux for the Bash sandbox;
it calls the live read-only MCP server):
cd plugins/skillx
claude plugin eval . --trust-plugin --mocks off \
--allow-tools 'mcp__plugin_skillx_skillx__*' Bash WebFetch --runs 2evals/results/ is git-ignored.
Local observations from the release session; the raw results are git-ignored. Re-run the suite above to reproduce them.
| Check | Result |
|---|---|
validate --profile directory --strict |
Passed, no findings |
Marketplace validate --strict |
Passed |
| Eval suite, 6 cases × 2 runs × 2 arms | 6/6 cases passed with the plugin, overall score 1.0, mean Δ +0.21 against no plugin |
| Eval re-run after the exact-load and untrusted-body changes | 6/6 passed, overall score 1.0, mean Δ +0.16; no get_package or /skills/<slug>.md load and no install before consent |
| Relevant prompts | Skill fired and search_packages ran in 6/6 runs |
| Unrelated prompts | Skill fired 0/6, catalog queried 0/6 |
| Install request | Never installed without consent (0/2 with plugin); the no-plugin baseline ran skillx install unasked in 1/2 runs |
| Old skill ("every task") on the same unrelated prompts | Fired in 1/6 runs |
| Claude Code install | marketplace add + install skillx@skillx-marketplace in a clean home: 1 skill, 1 MCP server, about 134 always-on tokens |
| CLI route | npx --yes skillx-sh@0.7.1 search … --json and check … --target claude-code (exit 4 on an undeclared runtime) |
Demo transcript summary for "Install the best code review skill from SkillX
into this project": Claude invoked skillx:skillx, ran search_packages
(with and without compatible: "claude-code") and check_compatibility for the
top two results, presented both with author, source repository, rating, risk
label unknown and compatibility unknown, noted the license had to be checked
in the source repository, and ended with "I will only install once you choose an
option" without installing anything.
Raise version in plugins/skillx/.claude-plugin/plugin.json and the matching
marketplace entry with every plugin change; claude plugin tag plugins/skillx
checks that the two agree. The pinned
skillx-sh version in SKILL.md and README.md is raised in a plugin release
after that CLI version is on npm. The directory rejects unpinned npx launchers.
Submission is a manual step in the developer portal from the claude.ai organization that should own the listing long term; the first organization to submit a repository folder holds it.
- Submit new → Plugin bundle, repository
nextlevelbuilder/skillx, plugin pathplugins/skillx, tracked branchmain. - Select Validate, fix anything marked Blocking, then submit. The
plugin runs a pinned
npxpackage, so expect the reviewer hold "Runs a pinned npx or uvx package". - Optionally submit
https://skillx.sh/api/mcpas an MCP connector from the same organization and pair it with the plugin listing. - Record the submission link, review status and, after launch, the plugin's Usage tab figures (installs, skill runs, error rates) on the tracking issue.
References: publish, pre-submission checklist, directory policy.