SkillBox
macOS · open source · local-first
Agent skills pile up across Codex, Claude Code, Cursor, and project folders until nobody
knows which copy is the real one. SkillBox collects every SKILL.md into one
managed library on your Mac, deploys it back out on purpose, and shows you what changes
before it writes.
~/.skillbox. Uninstall an agent and your library stays.
v0.9.5 · 75-second overview
The demo collects scattered skills into one managed library, then follows it through the Dashboard, a skill detail where one edit reaches every connected runtime, runtime-aware workspaces, review-before-write imports, and evidence-aware Calls with transparent local coverage.
~/.skillbox remains the local source of truth. Usage figures and deployment targets in the video are demo fixtures.The library
User skills and GitHub-backed remote skills sit in one managed store. Search them, filter by type or tag, see which are deployed where, and spot available updates without remembering which agent folder holds which copy.
Deployment
A deployed skill is an ownership-checked symlink, so the managed copy stays the real one. Open a skill to see every workspace it reaches, its version history, and its type, then change any of it from that single screen instead of editing four folders by hand.
Workspaces
Workspaces identify Agents, Codex, Claude Code, Cursor, and exact custom roots, global or
project-local. Each one reports its own skill count and usage, and
~/.skillbox stays the source of truth behind all of them.
Review before write
Installing from GitHub fetches one bounded repository ref and reviews it as a single snapshot. You see the resolved SHA, each child's status, and runtime compatibility before anything is written. The collection header applies one explicit User/Remote choice to the actionable children, then selects or clears the eligible set in one move. Mixed or unresolved type state stays blocked, every child remains independently managed, nothing deploys automatically, and collection-level update and rollback remain Phase D work.
Phase C first shipped in v0.9.0; the shared header controls shown here build on that reviewed per-child boundary. Validated copied skills may also group by normalized GitHub source as display-only provenance, without inventing branch, HEAD, or update authority.
Usage
Calls combine locally confirmed and defensible inferred invocations, and a year of daily activity is one heat map. History references stay separate because a mention is not a run, and the coverage panel shows exactly which local sources produced each number.
Audit trail
Sync local Codex, Claude Code, and Cursor histories to recover usage evidence without copying chat bodies. The History view keeps each evidence class and every management operation separate and readable, without claiming account analytics.
Trust boundaries
Runtime folders, GitHub URLs, downloaded archives, and existing skills are treated as untrusted input.
~/.skillbox holds managed copies, versions, metadata, and backups.
FAQ
Short answers for developers evaluating SkillBox as their local-first skill manager.
SkillBox is a local-first macOS app and CLI for organizing AI agent skills from one managed library. It focuses on Codex skills, Claude skills, and other SKILL.md-based workflows that need safer import, update, and deployment controls.
Managed user skills and GitHub-backed remote skills live under ~/.skillbox. Runtime
folders become deployment targets rather than the only place a skill exists, so removing an agent
never takes your library with it.
SkillBox targets local SKILL.md runtime folders such as .codex/skills,
.agents/skills, .claude/skills, and project-local skill roots. Workspace
targets are explicit, so one library can deploy to the runtimes you actually use.
GitHub-backed remote skills stay linked to their source and version history. SkillBox previews remote diffs before applying changes, so imports and updates can be reviewed before they touch runtime folders, and any version can be rolled back to.
A repository or tree ref is fetched once and reviewed as one snapshot. One collection-header User/Remote decision resolves actionable children, and Select all changes only the eligible set; unresolved, invalid, imported, system, or conflicted children stay unchanged. Each selected child remains an independent managed skill, and collection-level update and rollback are not included yet.
macOS is the current release target. SkillBox is open source, but the packaged desktop app and signed releases are currently focused on macOS.
Install
SkillBox currently targets macOS. Download the latest signed DMG or install the cask from the project tap.
brew tap santosli/tap
brew install --cask skillbox