New Tricks for VS Code
Teach your agents new tricks. The design-time workbench for agent skills: author, customize and test skills in a source repo, for Claude Code, Codex, GitHub Copilot and Cursor.
- Discover: one search across catalogs (skill repositories, Claude/APM marketplaces, skills.sh, Tessl, ClawHub and GitHub), with trust, licence and risk facets. Identical copies are grouped.
- Preview: read any skill (and its supporting files) without cloning it. Nothing is ever executed.
- Try: link an upstream skill into a project to see how an agent uses it, without vendoring it.
- Customize: vendor an upstream skill into your source repo, edit it, and keep merging upstream improvements in VS Code's three-way merge editor.
- Link & experiment: link your source repo's skills for your agents at user scope (edits are live), start an experiment on its own branch and worktree, link it into any project (git status stays clean), commit it, then merge it back or open a pull request.
- Lint: Agent Skills spec, structure, triggering quality and safety checks in the Problems panel.
- Publish: one pre-flight view, then a distribution repository installable by APM,
npx skills, Claude plugin marketplaces, Copilot, Codex and Cursor.
The extension is a thin client over the tricks CLI (tricks serve --stdio). Platform builds bundle the binary; otherwise install the CLI and make sure tricks is on your PATH, or set tricks.path.
Credentials are borrowed, never stored: GITHUB_TOKEN, then gh auth token, then your VS Code GitHub session (held in memory).
| |