The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Skill Builder MCP listing page.
Publish and install AI skill prompts from the command line — or author them in the web editor with help from an AI drafting assistant.
Results print as a dense table right in your terminal:
Filter by category, author, type, tag, and sort order:
Wire up shell completion (skills, tags, and authors complete as you type):
Downloads dialogue-flow.md and dialogue-flow.json to the current directory. Use --output ./my-skills to change where files land.
Meta skills install their dependencies automatically. Installing a meta skill
resolves its full dependency tree (deduped, cycle-safe) and installs every
required skill alongside it. Use --no-deps to install just the meta skill.
You'll be prompted for a password. Or set SKILL_PASSWORD to skip the prompt:
Your auth token is printed on success. Set it as SKILL_TOKEN so subsequent commands can use it:
(Or pass --token $SKILL_TOKEN to each command.)
From a markdown file:
The CLI reads the filename as the skill id and the first heading as the name. You can also use a JSON manifest:
A skill can declare what a model must be able to do to invoke it. The declaration lives on the spec, so it travels with the skill through the registry, the CLI, and MCP — an agent can tell whether it is able to run a skill before it spends a turn on it.
required capabilities gate execution: the registry refuses to run a skill on a
runtime that cannot provide one, unless the caller explicitly forces a degraded
run. preferred capabilities only weaken the result. Catalog ids:
Custom ids are accepted for anything the catalog does not model yet. info and
list show the contract, and install rolls it up across a meta skill's whole
dependency tree — a meta skill only runs where every skill it installs can run.
Creates a copy under your account that you can republish.
scripts/backfill-dependencies.mjs finds existing skills that implicitly
orchestrate other skills and proposes explicit dependencies + type: "meta"
for them. It reads every skill, detects references to other real registry
skills (by full id, install directives, or distinctive names), scores a
confidence level, and merges a curated set of known-correct mappings.
It is dry-run by default — it prints a report and writes nothing:
To apply, pass --apply with a token. Only skills owned by the token's account
are updated (others are listed and skipped), so run it once per publishing
account:
Confidence tiers: curated and high (explicit id / install / distinctive-name
references) apply by default; medium (fuzzy name match) only with
--include-medium; meta-no-deps (orchestration language but no concrete
dependency) is reported for manual review and never auto-applied.
Applying requires the registry to run the worker build that understands
type/dependencies(it derivestypeand regenerates markdown on save). The dry-run report works against any deployed version.
Coding agents — Claude Code, Cursor, Codex, and other MCP clients — can search the
registry and auto-install skills (resolving meta-skill dependency trees) via the
@dmzagent/skill-builder-mcp MCP server.
Tools: skill_search, skill_info, skill_suggest, skill_taxonomy, and
skill_install (writes skill files for the detected tool). See
packages/mcp/README.md for per-client setup.
Visit the skill studio at https://skills.dmzagent.com. The workspace is three
collapsible panes plus a settings drawer:
{{variables}} highlighted, examples and tests, and
a publish-readiness checklist that links into the field behind each gap.MIT