awesome-openclaw-skills: a markdown list of 5,400+ skills marked curated, not audited

Three different skill totals live in the same repository

The headline numbers do not agree, and the gap is large enough to matter when you are estimating coverage. The repository description says 5,400+ skills filtered and categorized from the official OpenClaw Skills Registry, while the header block above the title says 5300+ community-built skills organized by category. The table of contents gives a third number, and a much smaller one: adding up the eleven categories whose counts are visible produces 3,347, made up of Coding Agents and IDEs at 1,184, Web and Frontend Development at 920, Browser and Automation at 323, Productivity and Tasks at 207, AI and LLMs at 176, Git and GitHub at 167, Communication at 146, Marketing and Sales at 108, Speech and Transcription at 47, Smart Home and IoT at 41, and Data and Analytics at 28. The last row of that table is cut off mid-entry, so more categories exist and the visible sum is a floor rather than a total. Nobody is lying, but a reader who quotes a figure from this project has to say which figure they mean.

The lookup priority names a Local tier that the install table never defines

There are three documented ways to get a skill onto a machine, and they differ in how much they assume. The OpenClaw CLI path is a single command:

openclaw skills install <skill-slug>

The ClawHub CLI is the alternative for registry-managed skill folders that sit outside a full OpenClaw workspace:

npx clawhub install <skill-slug>

Manual installation means copying the skill folder yourself into `~/.openclaw/skills/` for global scope or `<project>/skills/` for workspace scope. Then one line settles conflicts between copies: Priority: Workspace > Local > Bundled. The problem is the vocabulary. The install table defines Global and Workspace, while the priority line refers to Workspace, Local, and Bundled, and nothing in the page says whether `~/.openclaw/skills/` counts as Local or as something else. A reader with a skill in two locations therefore has one documented winner and one tier whose name does not match any path they were given.

Internal links still use the old awesome-clawdbot-skills repository name

The repository is named awesome-openclaw-skills, but several of its own links are not. The pull request badge points at a query against VoltAgent/awesome-clawdbot-skills rather than the repository you cloned, and the security notice asks you to open an issue at github.com/VoltAgent/awesome-clawdbot-skills/issues. Both paths belong to the earlier name of the project, which is consistent with the rename that turned a clawdbot list into an OpenClaw list. The practical effect is on the two workflows that matter most for a directory like this. Someone following the pull request link to see what changed recently, and someone reporting a skill they believe is unsafe, both leave the repository they are reading. The submission instructions are also explicit that personal repositories, gists, and other external sources are refused, so the ClawHub registry is the only entry point, and that registry link is the one route through the project that is not affected by the naming mismatch.

The malicious filter excluded VirusTotal, the same source the security notice tells you to use

The exclusion table is the most useful part of the repository and the easiest to misread. Five rules are applied, and the numbers sum exactly to the stated total of 7,215 skills not taken from the official registry: 4,065 dropped as possibly spam, meaning bulk accounts, bot accounts, and test or junk entries, 1,040 as duplicates or similar names, 851 as low quality or non English descriptions, 886 in crypto, blockchain, finance, or trade, and 373 as malicious, identified by security audits published by researchers, explicitly excluding VirusTotal. Then the disclaimer lands: skills in this list are curated, not audited, and they may be updated, modified, or replaced by their original maintainers at any time after being added. Read together, those two passages mean a listing is evidence of a link that existed and a name that was not duplicated. It is not a security check, and 373 known-bad entries removed tells you nothing about the ones that arrived after the audit was published.

Category depth is lopsided, so browsing quality depends on which bucket you need

The taxonomy is the real product here, and its shape decides what a reader finds. Two categories carry more than half of the visible entries between them, with Coding Agents and IDEs at 1,184 and Web and Frontend Development at 920, followed by a long tail: Browser and Automation 323, Productivity and Tasks 207, AI and LLMs 176, Git and GitHub 167, Communication 146, Marketing and Sales 108, Speech and Transcription 47, Smart Home and IoT 41, and Data and Analytics 28. Two things follow. A developer looking for a code review or editor integration has a list long enough that narrowing it is their problem, while someone looking for analytics tooling has 28 candidates and someone looking for smart home control has 41. And because the buckets are editorial rather than keyword derived, the 1,040 duplicates removed by similar name do not mean the largest bucket is clean, it means the surviving overlaps are ones the curators judged different enough to keep.

The repository is markdown under categories/, and the searchable site is built outside it

The tree is small and entirely editorial: .claude/, .github/, CONTRIBUTING.md, LICENSE, README.md, and categories/, with no source directory, no package manifest, and no primary language detected. That means the list is the artefact, and the way to read it is by hand or by scraping. Two consequences follow for anyone building tooling on top of it. There is no machine-readable index in the repository, no schema for what a category file contains, and no published data structure, so a script that wants the list has to parse the markdown and re-derive it whenever the categories change. And the repository has no GitHub releases, so there is no snapshot to pin: a category count can change with a single commit and no release note, and a downstream scraper has no way to tell a deletion from a temporary edit. The searchable site at clawskills.sh is explicitly managed separately by the maintainers, so the site and the repository can disagree.

Three paid blocks sit between the install instructions and the security notice

The page is not only a list. After the installation section it runs a series of sponsored blocks for ecosystem tools, and each link carries campaign parameters. Crawlbase is presented as web data infrastructure with JavaScript rendering, proxy rotation, and anti bot handling, plus an MCP server exposing crawl, crawl_markdown, and crawl_screenshot, and it claims to be trusted by 70,000+ developers. Cloudways offers managed hosting for OpenClaw on dedicated infrastructure with managed updates, backups, SSL, and security controls, and hands out $10 in credit with the promo code VOLTAGENT. SerpApi offers real time Google Search, YouTube, Amazon Product, and web search data through one API. The surrounding prose about agents needing proxy rotation, CAPTCHAs, and HTML parsing handled for them is sponsor copy, not a curation finding, and the block inviting other tools to be featured is part of the same arrangement. Readers should treat that region as advertising and the category tables as the content.

Only skills already published on ClawHub can be listed, and the site lags the repository

The submission rules are strict in a way that shapes what the list can ever contain. A skill qualifies only if it is already published on ClawHub, which is OpenClaw's public skills registry, and links to personal repositories, gists, or any other external source are refused outright. Contributors are asked to put the ClawHub link for their skill in the pull request description, using a form like https://clawhub.ai/steipete/slack, and CONTRIBUTING.md holds the detail. The same paragraph adds that the listings on the site are handled separately by the maintainers, which splits the publication path in two: a skill can be in the repository before it appears at clawskills.sh, or in one and not the other. The result is a directory that cannot showcase unpublished work, cannot link a maintainer's own repository, and never presents a single canonical index, which is a sensible rule for a security conscious list and a real inconvenience for anyone automating against it.

Editorial conclusion

Treat awesome-openclaw-skills as a map rather than an endorsement: it is a good way to find out which skills exist in a category, and a bad basis for deciding that any of them are safe. Use it if you run OpenClaw locally and want a shortlist to review yourself, and expect to do the review, because inclusion means a link was published on ClawHub and passed a name and description filter. Before installing anything from it, open the skill's page on ClawHub and read the VirusTotal report, read the source, and check whether the skill's maintainer can still change it after you install. If your use case is analytics, smart home, or speech, look elsewhere first: those categories hold 28, 41, and 47 entries while Coding Agents and IDEs alone holds 1,184.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository