Skip to main content
This plugin puts the DecimalAI skills registry inside Claude Code as three slash commands. Search shows each skill’s evidence next to its install line: how much it beat a run with no skill on a benchmark (its “lift”), and its safety band.

Install

No account, no API key, nothing to pip install. It has no hooks and no MCP server, and nothing runs until you type a command. Search and install run a small Python script, so they need python3 and curl.

The three commands

/decimalai:search <query>

Searches the public registry and prints up to ten results, with each skill’s test results. Add --category <cat> or --sort recommended|lift|popular|top_rated to narrow or reorder them. Each row gives the exact /decimalai:install line, the folder it lands in, a description, the SkillScore, the lift in points, the safety band and how many agents have the skill now. A score of New or a blank lift means it hasn’t been measured. Searching for sql injection lists detecting-sql-injection-via-waf-logs at +54.55 points, safety passed.

/decimalai:install <url_slug>

Writes the skill’s folder exactly as the registry renders it into .claude/skills/<install_name>/, where Claude Code reads project skills. That’s SKILL.md plus its scripts/, references/, templates/ and assets/. The folder is the skill’s install name, so a namespaced owner/skill lands in .claude/skills/owner-skill/.
  1. It reads the skill’s record and prints its safety band and benchmark line. A blocked skill stops here. Any other band except passed waits for your yes before anything is fetched.
  2. It fetches the folder from GET /api/v1/registry/skills/<id>/bundle: every file with its exact text, SHA-256 and mode.
  3. Before writing anything, it refuses:
    • a folder name loaders don’t accept
    • a path outside those four folders, including .., absolute paths and backslashes
    • a second SKILL.md anywhere in the tree
    • a mode other than 0644 or 0755
    • any file whose bytes don’t hash to the SHA-256 the registry declared
  4. It writes into a staging folder, reads each file back to re-check its hash and mode, and only then moves the folder into place.
If /<install_name> isn’t available in your session, restart Claude Code. Pass the url_slug that search printed, or the skill’s id, not its name: a registry name can be owner/skill, and a / can’t be part of a folder name. The output can carry two kinds of note:
  • not written: is a binary attachment, which the registry doesn’t store, or a nested SKILL.md, which would load as a second skill.
  • registry note: includes any allowed-tools entry the registry dropped. A skill your organization didn’t author may pre-approve only Read, Grep and Glob, so Claude still asks before running anything else.
A folder that already holds exactly these bytes is left alone. One that differs is replaced only after you say yes, and it’s moved to .claude/decimalai-backup/ first. A symlink or a file in the way is always refused. The copy is unmanaged: nothing records it or updates it. The file carries the registry’s identity stamp, so you can later tell what you installed and whether you’re behind. For a pinned install, see Managed skills.

/decimalai:check

Writes the two files behind the regression-check Action, with your repo’s real agent name filled in:
  • scripts/init_for_decimal.py, which registers what your agent is built from (its manifest) on each pull request
  • .github/workflows/decimal-regression-check.yml, the workflow itself
The workflow needs a DECIMAL_API_KEY repo secret because it checks changes against your own production runs. Adding that secret is up to you.

The command that refuses

The registry indexes skills scraped from public repositories, and some are hostile. Search for sql injection and one match is performing-second-order-sql-injection. Its pattern scan came back clean: no secrets, no reverse shell, no step that sends data out. The intent review, which reads the file for its purpose, rejected it anyway for giving “actionable instructions” for attacking a database. That’s why the two checks are separate. /decimalai:search prints it last, apart from the results, with no install line:
/decimalai:install refuses it before fetching anything, and the registry won’t serve a blocked skill’s files on any route.
The ask-first step for caution and unreviewed skills is a prompt in your own session, not a boundary the model can’t cross. Treat it as a convenience. The check is the band on the skill’s page: how skills are vetted.

Other ways into Claude Code

All three ways write the registry’s exact bytes. They differ in what keeps track of the copy.

The one-line copy

Pick Claude Code in the Use this skill card on a skill’s page, open Other ways to add it and choose Just the file. It shows one command with the skill’s real install name filled in:
Claude Code picks the skill up in its next session. It writes SKILL.md alone, so it suits skills with no bundled files. Copy SKILL.md, at the top of the page, copies the same file. More: Raw URLs.

Managed skills (decimalai skills add)

For skills a project keeps pinned, updated and reproduced in CI, use the CLI. Public skills need no API key:
It writes .agents/skills/<install_name>/ and records the exact bytes in .decimal/skills.lock. Commit .decimal/ and .agents/. Claude Code doesn’t read .agents/skills/ as project skills. It loads that folder’s parent as a plugin. Start it with the repository’s absolute path, because a relative --plugin-dir silently loads nothing from a subdirectory:
The project’s plugin is named decimalai-skills, so its skills appear as /decimalai-skills:<install_name>, and it loads side by side with this marketplace plugin. Keep a skill in one place: a second copy in .claude/skills/<install_name>/ loads it twice, and /decimalai:install warns you when it sees both.
In a project connected to DecimalAI (it has a .decimal/link.toml), skills add adds the skill to the agent there and writes no files, so Claude Code won’t see it. Use /decimalai:install for Claude Code’s copy.
Skills on disk covers the lock, CI and updates. The CLI page has every command.

What leaves your machine

The search query and the skill’s url_slug or id, sent as anonymous GETs to api.decimal.ai with the user agent decimalai-claude-plugin/<version>. Nothing is stored against you. The requests appear in server logs, as described in §1.4 of the privacy policy. The plugin sends no telemetry of its own. The folder request is labelled channel=plugin, and the registry adds it to that skill’s fetch count for the day, with nothing that identifies you. Fetch counts are kept apart from a skill’s measured effectiveness, so a local install never moves those numbers.

Not using Claude Code?

/decimalai:install writes only to .claude/skills/. For Cursor, Copilot, Cline and other tools, see Cursor and other tools.

Source

decimal-labs/decimalai-claude-plugins, MIT. You can read each command’s prompt and the one Python script they run before you install it.