Guides

Installing items

dex install REF... --target claude|copilot|dir [--scope user|project] [--path DIR] [--allow-exec]

dex install copies items out of the installed package into a tool’s native layout. It never runs anything. You can pass several refs at once, and plugins expand to their members.

Targets

`--target claude` (default)

Kind	Destination	Notes
skill	`<base>/skills/<name>/`	whole folder, minus Dex’s `ananke.yaml`/`dex.json`; licence files included
agent	`<base>/agents/<name>.md`	entry file + `<name>.md.LICENSE`
prompt	`<base>/commands/<name>.md`	becomes a slash command
plugin	members, as above
instruction, hook, workflow	—	reported as “no native claude layout”; use `--target dir`

<base> is ~/.claude with --scope user (default) or ./.claude with --scope project. --path overrides it.

`--target copilot`

Kind	Destination
skill	`.github/skills/<name>/`
agent	`.github/agents/<name>.agent.md`
instruction	`.github/instructions/<name>.instructions.md`
prompt	`.github/prompts/<name>.prompt.md`
hook	`.github/hooks/<name>/` (needs `--allow-exec`)
workflow	`.github/workflows/<name>.md`
plugin	members, as above

The base is ./.github in the current directory; --path overrides it.

`--target dir --path DIR`

Copies the complete normalised item to DIR/<kind-plural>/<namespace>.<name>/, including ananke.yaml and dex.json. Use it to vendor items into a repo, feed another harness, or review an item before installing it.

Names on disk

Native targets use the plain item name. Skills require the folder name to match name: in SKILL.md. If you install two items with the same name from different sources, the second overwrites the first. Check with dex search <name> first.

Licences travel with the item

Folder items keep their own licence files and any LICENSE.upstream / NOTICE.upstream.
Single-file items (agents, instructions, prompts, workflows) get a sibling <file>.LICENSE when the governing text came from outside the file.

Keep these files if you redistribute the installed items.

Executable items

Items with scripts are tagged executes-code. They install normally, but read the scripts before an agent runs them.
Hooks run automatically on session events, so they are refused unless you pass --allow-exec:

dex install awesome-copilot/secrets-scanner --target copilot
# – hook:awesome-copilot/secrets-scanner: hook runs code — pass --allow-exec after reviewing it
dex install awesome-copilot/secrets-scanner --target copilot --allow-exec

Updating and removing

Upgrade the package (pip install -U dex-trove) and run the same dex install again; destinations are replaced. To remove an item, delete its folder or file. Dex keeps no install state of its own.

Adapting across tools

Upstream items are written for a particular tool. When moving between tools:

Agent tools: lists use the source tool’s names (Copilot’s codebase, editFiles, …). Claude Code ignores tool names it doesn’t know. Adjust the frontmatter if the agent relies on specific tools.
Skills follow the cross-tool Agent Skills spec and are the most portable kind.
Instructions have no Claude Code equivalent. Reference the file from CLAUDE.md or wrap it in a skill.

Edit this page on GitHub ↗