Questions

Three routes, and they work differently. subagents.sh publishes a CLI: "Install any subagent with a single command: npx @augmnt-sh/subagents add owner/repo/subagent-name", which "downloads the subagent to .claude/agents/ in your project directory", with list, remove and update commands beside it. aitmpl.com is a browse-and-basket catalogue — "browse components and click the + button to add them to your stack for easy installation" — covering agents alongside skills, commands, settings, hooks, MCPs and plugins. And any subagent published in a public GitHub repository can be copied into .claude/agents/ by hand, which is what both of those are automating.

Also answered on /best-subagents

It depends entirely on what is being counted, and the published numbers are not comparable. subagents.sh prints "40 subagents" on its own header — a deliberately narrow catalogue. aitmpl.com's headline is "1000+ Agents, Commands, Skills & MCP Integrations" across all seven artifact kinds, of which its agents section held 421 when browsed on 14 August 2026. SkillWorks counts every subagent it can find published in a public repository, which is a much larger number and a much worse one to quote unqualified — because the interesting figure is not how many exist, it is what share of them load.

Also answered on /best-subagents

None of the ones compared here publishes that it does. subagents.sh documents the file format a subagent must follow — name required, description and tools optional — but states nothing about validating listed files against it. aitmpl.com's front page publishes no validation, parse check or per-component health signal. Where a check exists elsewhere in this market it is explicitly editorial rather than mechanical: awesome-claude-code's maintainer writes that "the bot is a mechanical form-validator, it does not review the quality of the recommended resource". A directory that has never opened the file cannot tell you the file is broken, which is the specific gap SkillWorks exists to fill.

Also answered on /best-subagents

Because a subagent is a markdown file whose behaviour depends on YAML frontmatter, and YAML that a human reads as fine is not the same as YAML a parser accepts. subagents.sh publishes the contract itself — "name: yes, human-readable name for the subagent" is the only required field, with description and tools optional — and the common failures are exactly the ones that contract implies: frontmatter that does not parse, a missing name, a missing description. Those are invisible in a catalogue listing, in a stars count and in an install count, because none of those ever opens the file. SkillWorks re-reads every file it indexes, every night, and scores it on whether it loads.

Also answered on /best-subagents

Popularity is an input to it and not the whole of it. The bar on this page is a published one — a score of 78 or better, at least 1,000 stars, and activity within 45 days — and the ranking then takes the highest-scoring qualifying subagent from each repository rather than the top N overall. That last rule is the important one: every artifact in a monorepo inherits the same maintenance, adoption and trust components, so a naive top-N degenerates into a listing of whichever repository publishes the most files. The other lists in this market order by install count (subagents.sh), by "Most Popular" or alphabetically (aitmpl.com), or not at all — awesome-claude-code and Smithery's skills catalogue publish no ordering.

Also answered on /best-subagents

Four routes, and they overlap less than they look. aitmpl.com lists plugins alongside six other artifact kinds and installs them through a stack you add to. Smithery is an MCP-first registry with a skills catalogue beside it, oriented around hosted connections — "Add to toolbox" rather than a file you copy. awesome-claude-code is a curated GitHub list whose scope explicitly includes plugins — "and also we have plugins" — alongside skills, agents, status lines and tooling. And any repository publishing a marketplace.json is a marketplace Claude Code can add directly, which is the mechanism the other three sit on top of and the one SkillWorks indexes.

Also answered on /best-plugins

Only two of the three publish a route at all. awesome-claude-code publishes the strictest and the clearest: recommendations go through a web issue form, one resource at a time, and "any resource that is recommended must either: (i) be at least 14 days old … AND show signs of active development … OR (ii) have at least 100 stars", with failures "closed automatically". Smithery publishes self-serve publishing as a first-class route, with its cost addressed in the pricing FAQ. aitmpl.com's own front page publishes no submission route; what it carries beside its featured slots is "Promote your component →", which is paid placement rather than a listing process.

Also answered on /best-plugins

Not in a way any of them publishes. aitmpl.com's front page carries no validation, parse check or per-component health signal. Smithery offers a "Verified" filter on its skills catalogue but does not publish what verification tests, so what is visible is the presence of a check rather than a description of one. awesome-claude-code is explicit that its automation does not do this — "the bot is a mechanical form-validator, it does not review the quality of the recommended resource" — and that the human review is best-effort. SkillWorks re-reads every manifest it indexes every night and publishes a 0–100 score for whether it loads, which is the entire reason this index exists.

Also answered on /best-plugins

The published figures are not measuring the same thing, so quoting one against another is misleading. aitmpl.com's headline is "1000+" across all seven artifact kinds, of which its plugins section held 33 when browsed on 14 August 2026. Smithery's catalogue reported "18,965 skills found" the same day, alongside 715 MCP servers — but a Smithery skill is not a Claude Code plugin, and the overlap is partial. awesome-claude-code publishes no item count at all; its headline numbers are repository numbers. SkillWorks counts repositories publishing a marketplace.json, which is the definition Claude Code itself uses.

Also answered on /best-plugins

Because they fail in different materials. A plugin is a JSON manifest, and JSON either parses or it does not — the tooling that produced it would usually have caught the error first, and the observed failure rate is roughly two tenths of one percent. A subagent is prose behind YAML frontmatter, where a file that reads perfectly to a human can still be missing a required field or carrying frontmatter a parser rejects, and roughly one in five published subagents does not load. Averaging those two into one "how much of this ecosystem is broken" number would hide the only interesting thing about either, which is why SkillWorks scores and ranks them on separate pages.

Also answered on /best-plugins

A repository publishing a marketplace.json manifest, which Claude Code can add directly so its plugins become installable. That mechanism is the layer underneath every directory on this page — aitmpl.com, subagents.sh, Smithery and awesome-claude-code are all ways of finding things that ultimately live in a repository. This index takes the mechanism as the unit: it finds every repository publishing a marketplace.json, counts the plugins each carries, checks whether the manifest is valid, and re-runs that every night.

Also answered on /marketplaces

Use aitmpl.com for breadth — it is the only one carrying all seven artifact kinds under one search, with a stack you add to. Use subagents.sh for subagents specifically, because its CLI writes straight into .claude/agents/ and its ranking is install count from real CLI telemetry. Use Smithery when the thing you want is a connection to a service rather than a file, because it runs the auth and credentials for you and its catalogue is the largest by a wide margin. Read awesome-claude-code when you are still working out what exists, because it is the only hand-picked list and the only one covering essays, field guides, alternative clients and cost tooling that no installer can install.

Also answered on /marketplaces

None of the four publishes a load check. Smithery offers a "Verified" filter on its skills catalogue but does not publish what verification tests, so what a reader can confirm is that a check exists, not what it does. awesome-claude-code is the most candid about it: its maturity gate is mechanical and its quality review is not — "the bot is a mechanical form-validator, it does not review the quality of the recommended resource", and "recommendations are reviewed in a best-effort way". aitmpl.com and subagents.sh publish nothing on the subject on the pages read. Where a cell in the table above says "Not published", it means exactly that: the page was read and did not carry the figure, not that the check does not exist.

Also answered on /marketplaces

Three different answers and one absence, and the difference matters more than it sounds. subagents.sh sorts by install count gathered from its own CLI — "we use anonymous telemetry from CLI installations to track popularity" — which is the only ordering here that measures the artifact rather than its repository. aitmpl.com offers two sorts, "Most Popular" and alphabetical. Smithery's skills catalogue prints per-item install figures but publishes no sort order and offers no sort control. awesome-claude-code groups by category with no ordering inside a category, which is the correct behaviour for a curated list and useless as a ranking.

Also answered on /marketplaces

By crawling rather than by curation, and then by scoring. Every repository publishing a Claude Code marketplace.json is a candidate, every skill, subagent and plugin inside it is re-read, and each one gets a published 0–100 score for whether it actually loads. Nothing is hand-picked and nothing is paid for. That is deliberately the opposite trade-off from awesome-claude-code, which is hand-picked and says so, and it is why the two are worth reading together: a curated list tells you what a person thought was good, and an index tells you what is actually there and what state it is in.

Also answered on /marketplaces