Homebrew for AI coding skills.
Boost is a command-line package manager that searches, installs, and manages AI coding skills from GitHub-hosted registries — then wires them into Claude Code, Windsurf, and Cursor at once. One command replaces the copy-paste-and-pray dance of setting up AI agent skills by hand.
What is a “skill” — and why manage them?
AI coding agents get better when you give them reusable knowledge. Boost makes that knowledge installable.
A skill is a SKILL.md file — Markdown with a little YAML frontmatter — that teaches an AI agent a repeatable capability: a workflow, a set of coding rules, or domain knowledge. Agents like Claude Code load them automatically from ~/.claude/skills/.
The problem: finding good skills, installing them into every agent, keeping them updated, and sharing them with your team is all manual copy-paste. Boost is the package manager that automates it — the way Homebrew handles Mac apps or npm handles JavaScript. It also manages rules, workflows, and configs under the same roof.
--- name: brainstorming description: Structured ideation & divergent-thinking facilitation version: 1.4.0 --- # Brainstorming When the user wants to explore ideas, facilitate structured divergent thinking: 1. Diverge — generate widely, defer judgment 2. Cluster — group related concepts 3. Converge — score & select
From GitHub to every agent
Boost clones registries, builds a searchable cache, installs skills to one canonical store, then symlinks them into each AI agent — all tracked by a lock file.
GitHub registries
Pre-tapped public registries + an 8,400-skill discovery index
~/.boost/repos/
Shallow git clones
~/.boost/cache/
JSON catalogs from SKILL.md frontmatter
~/.agents/skills/
Single source of truth for installed skills
.skill-lock.json
v3 schema — pins version, commit & content hash
Claude Code
~/.claude/skills/
Windsurf
~/.windsurf/skills/
Cursor
~/.cursor/skills/
Search → install → done
The thin bash shim finds your Python and runs with zero dependencies. After that, the core loop is three commands.
$ boost search jira ▰▰▰▰ jira-integration Sync commits and PRs to Jira tickets ★ curated ▰▰▱▱ commit-messages Conventional, atomic commit message discipline 2 matches · ranked by Claude Haiku relevance $ boost install jira-integration ✓ copied to ~/.agents/skills/jira-integration ✓ linked → claude-code · windsurf · cursor ✓ lock updated (.skill-lock.json) ╭─ installed ────────────────────────────────────────╮ │ Installed 1 new skill; quality score 92/100 │ │ jira-integration → claude-code · windsurf · cursor │ │ next: boost info jira-integration │ ╰────────────────────────────────────────────────────╯ $ boost doctor ==> boost doctor ✓ 14 skills installed · 2 taps synced · 0 broken links ✓ lock file integrity OK · log rotation healthy ● healthy # share the whole setup with your team: $ boost bundle dump > Boostfile # teammates run: boost bundle install
73 commands, 8 groups
From basic install/update to AI-powered authoring, governance, and team rollouts. Filter live to find what you need.
No commands match — try install, audit, or clear the filter.
The capabilities that set it apart
Boost treats skills as first-class, governed, shareable artifacts — not just files you copy.
AI-powered authoring
Claude ranks your searches by relevance, summarizes skills, drafts new ones from your codebase or chat history, and previews behavior before you install.
Governance & safety
Enforce org policies, detect drift from source, block unsafe skills against a safety list, and verify integrity with content fingerprints and install attestation (TOFU).
Team & collaboration
Roll skills out to cohorts with A/B testing, switch context with profiles, see a team activity feed, and find who has expertise. Share exact setups via Boostfile bundles.
Multi-agent & integration
Install once, land in Claude Code, Windsurf, and Cursor. Run as a
native MCP server, register a boost:// one-click
protocol, serve a catalog over HTTP, or auto-sync on a schedule.
Reproducible & recoverable
A v3 lock file pins exact state. Preview diffs before syncing, snapshot and restore whole environments, replay history, and pin versions so nothing changes underfoot.
Self-healing & secure
Auto-repair broken symlinks and stale caches, detect unused or conflicting skills, and stay hardened: sandbox-safe paths, safety-pattern audits, and SHA-verified state.
Pre-tapped registries + open discovery
Boost ships pre-tapped to curated public skills. Beyond that,
boost discover indexes 8,400+ skills across GitHub.
Tap your own team’s repo anytime.
| Tap | Repository | Skills | Focus |
|---|---|---|---|
| anthropics/skills | github.com/anthropics/skills | — | Official Anthropic skills — documents, artifacts, MCP builders, agent workflows |
| obra/superpowers | github.com/obra/superpowers | — | Community powerhouse — TDD, debugging, planning, subagent workflows |
| your-org/skills | any repo with SKILL.md files | ∞ | Your team’s registry — boost tap your-org/skills (private repos work over ssh) |
Skill counts measured from the live repositories ·
build the GitHub-wide index with boost index, browse it with boost discover
One shim, zero setup
Copy a single bash shim onto your PATH. It finds Python 3.9+
and runs — no PyPI, no Artifactory, no manual setup.
# clone, link the shim onto your PATH, verify git clone https://github.com/your-org/boost ~/.boost-src ln -s ~/.boost-src/boost ~/bin/boost boost --version # add the recommended registries and go boost tap --defaults boost search brainstorming