sx

The 30-second version

1. Download the app (sx-app-* for macOS, Windows, Linux) and open it — it asks one question and sets up your library.
2. Point it at a folder your team already syncs — Dropbox, Google Drive, iCloud, OneDrive. That folder is the team vault. No server, no git, no accounts.
3. Drag a skill in. It's published. Teammates who point at the same folder get it on their machine, in every AI tool they use.

Command-line person? The same thing, three lines (npm-style manifest + lock under the hood):

brew install sleuth-io/tap/sx
sx init --type path --path ~/Dropbox/sx-vault
sx add ~/.claude/skills/my-skill

Why sx?

AI assets — skills, MCPs, agents, rules, commands, hooks — usually live inside a single Git repo. The moment you want them in another repo you copy-paste, and they drift out of sync with no source of truth. sx manages complex sharing and distribution of AI assets for real-world teams:

• Zero infrastructure — a folder your team already syncs is a full vault; no server, no accounts
• No Git knowledge required — non-technical teammates share skills too; the plumbing stays hidden
• Every AI tool, including the web ones — one asset installs into every major AI assistant
• Share across projects and teams — manage an asset once, install it into any repo or team
• The right assets, not all of them — scope to an org, repo, path, team, bot, or user; no context bloat
• Grows with you — versioning, adoption stats (sx stats), and an audit trail (sx audit) when you need them

App or command line — your choice

Two front doors, one library. The app and the CLI share a single configuration, so a library added in one shows up in the other — mixed teams never drift.

• The app — for everyone. Create and edit skills in a built-in editor (or drag files in), publish to your team, install into your AI tools with one click. No terminal anywhere. Download for macOS, Windows, or Linux (the sx-app-* artifacts).
• The CLI — for engineers and automation. The same operations, scriptable: pin it in CI, wire it into hooks, drive it from your own tools. Install below.

Quickstart

Install via Homebrew (macOS/Linux):

brew tap sleuth-io/tap
brew install sx

Or via shell script:

curl -fsSL https://raw.githubusercontent.com/sleuth-io/sx/main/install.sh | bash

Then initialize in your vault or project:

sx init

From here: add skills, share them with the right people, and see what's used.

Manage — capture, version, and observe

Add assets to your vault (sx auto-detects the type):

sx add /path/to/my-skill
sx add ~/.claude/skills/my-skill            # an existing Claude Code skill
sx add code-review@claude-plugins-official  # a plugin from a registry

Your AI assets stay exactly as they are — sx just wraps them with metadata for versioning and stores them in its vault format.

Multiple vaults? Use profiles to switch between them:

sx profile add work        # Add a new profile
sx profile use work        # Switch to it
sx profile list            # See all profiles

See what's actually used — track adoption and token usage across your team:

sx stats                   # adoption dashboard
sx stats --since 7d --json # machine-readable

Distribute — install to the right scope, on any client

# Install everything scoped to you, into the current project
sx install

Install targets — pick who sees which asset:

sx install my-skill --org                               # everyone in the vault
sx install my-skill --repo github.com/acme/infra        # only inside that repo
sx install my-skill --path github.com/acme/infra#docs/  # one path in a repo
sx install my-skill --team platform                     # every member of a team
sx install my-skill --user [email protected]               # a single user
sx install my-skill --bot python-backend                # a bot identity (CI runner, agent)

See docs/scoping.md for the full overview and
links to a per-scope doc for each install target.

Use your vault from claude.ai or chatgpt.com — expose it as an MCP
endpoint via the skills.new relay:

sx cloud connect       # opens skills.new, paste back the attach line
sx cloud serve         # keep this running — Ctrl+C exits
sx cloud status        # prints the MCP URL to paste into claude.ai / chatgpt.com

The relay forwards requests over a WebSocket — vault
content stays local. See docs/cloud-relay.md.

Govern — audit and migrate

Audit — every team and install mutation is recorded:

sx audit                                   # recent team/install mutations
sx audit --actor [email protected] --since 30d --event install.set

Migrate a whole vault (assets + versions, teams, bots, scopes, audit, usage):

sx vault copy --from skills-new --to git-vault   # preview (read-only)
sx vault copy --from skills-new --to git-vault --yes

See docs/copy.md for directionality and what's lossy. A gated change-request flow (RBAC) is on the roadmap.

What can you build and share?

• Skills - Custom prompts and behaviors for specific tasks
• Rules - Coding standards and guidelines that apply to specific file types or paths
• Agents - Autonomous AI agents with specific goals
• Commands - Slash commands for quick actions
• Hooks - Automation triggers for lifecycle events
• MCP Servers (experimental) - Model Context Protocol (MCP) servers for external integrations
• Plugins - Claude Code plugin bundles with commands, skills, and more

Portable agents, not just files

An AI agent is only as good as what it knows and what it's allowed to do. sx lets you describe that once — define a Bot and attach the agent's prompt plus the skills, rules, commands, hooks, and MCP servers it depends on — and install it unchanged across any client, coding or not.

• One definition, every tool — the same agent runs unchanged on every supported client, coding or not; sx writes each one's native format on install
• Bundle the whole capability — skills, rules, commands, hooks, and MCP config travel together as versioned assets, not loose files scattered across repos and machines
• Decoupled from any one vendor — AI tools are commoditizing; the agents running on them shouldn't be locked in. Describe them in a portable format you own and carry them between tools as the landscape shifts

Put them to work: sx + hetchy

sx gets your skills and agents onto every client your team uses. hetchy is where they run unattended: connect a repo, send a request from the web UI, Slack, or Linear, and hetchy runs a coding agent in an isolated sandbox, then opens a reviewable pull request with the evidence attached.

The two are wired together. Every hetchy run installs agent personas and scoped skills from an sx vault before the agent starts — its public vault by default, or your own Skills.new vault once an org admin adds an SX key. Publish here and it's in the next run:

sx add ~/.claude/skills/migration-reviewer   # publish to your vault
# hetchy installs it into the next sandbox — no copy-paste, no redeploy

hetchy embeds sx as a Go library (pkg/sxvault) rather than shelling out to the CLI, so vault access is part of the run itself.

hetchy is self-hostable and Apache-2.0 — docker compose up. See docs/hetchy.md for the end-to-end workflow.

Distribution models

Choose the right distribution model for your team:

Shared folder (Small teams, zero infrastructure)

Put the vault in a folder your team already syncs — Dropbox, Google Drive,
OneDrive, or iCloud. No git, no GitHub account, no server. See
docs/synced-folders.md.

sx init --type path --path ~/Dropbox/sx-vault

Local (Personal)

Perfect for easily sharing personal tools across multiple personal projects

sx init --type path --path my/vault/path

Git vault (Small teams)

Share assets through a shared git vault

sx init --type git --repo-url [email protected]:yourteam/skills.git

Skills.new (Large teams and enterprise)

Centralized management with a UI for discovery, creation, sharing, and usage analytics

sx init --type sleuth

Plugin marketplace (no sx required)

Every git or path vault is also a Claude Code / Codex plugin
marketplace — sx generates and maintains .claude-plugin/marketplace.json
and .codex-plugin/plugin.json on every publish. Teammates who don't run
sx can install the library's skills directly from their AI tool:

# Claude Code
/plugin marketplace add yourteam/skills
/plugin install skills@skills

# Codex
codex plugin marketplace add [email protected]:yourteam/skills.git

Each collection in the vault is also exposed as its own Claude Code
plugin, so people can install just the slice they need. Private repos
work through normal git credentials. Plugin installs deliver skills only
(rules and per-team scoping still need sx install) — see
docs/plugins-spec.md.

How it works

sx follows the manifest-and-lock pattern used by npm, cargo, and uv:

1. Manifest (sx.toml) — the vault's source of truth. Lists every
   managed asset, its install scopes (org, repo, path, team,
   bot, user), and team definitions (members, admins, repositories).
   Committed to git / path vaults. See docs/manifest-spec.md.
2. Lock file — a per-user resolved artifact. sx install reads the
   manifest, resolves team and user scopes against the caller's git
   identity, and writes the result to the user's cache directory
   (~/<cache>/sx/lockfiles/). When the resolved lock changes, the
   previous file is rotated with a timestamp so old installs stay
   reproducible.
3. Audit + usage streams — every team/install mutation appends an
   audit entry to .sx/audit/YYYY-MM.jsonl; usage events append to
   .sx/usage/YYYY-MM.jsonl. Query them with sx audit / sx stats.

High level: manage assets in one vault, distribute them
globally, per repo, per path, per team, per bot, or per user —
auto-installing on new Claude Code sessions so everyone stays in sync — and
govern every change through the audit and usage streams.

Use sx as a Go library

Everything the CLI does to a vault is also a package. Publish skills and
agents, manage bots and teams, and browse or download assets from your own
program against Skills.new, Git, or local Path vaults through one Client:

import "github.com/sleuth-io/sx/pkg/sxvault"

ctx := context.Background()
client, err := sxvault.OpenSkillsNew("https://app.skills.new", token)
if err != nil {
    log.Fatal(err)
}
if err := client.PutSkillZip(ctx, sxvault.SkillZipSpec{
    Name: "lint-helper", Version: "1.0.0", ZipData: zip,
}); err != nil {
    log.Fatal(err)
}

See docs/library.md for the full API guide.

Supported Clients

Client	Status	Notes
Claude Code	✅ Supported	Full support for all asset types
Cline	✅ Supported	Skills, rules, workflows as commands, MCP servers, hooks
Codex	✅ Supported	Skills, commands, agents, MCP servers
Cursor	✅ Supported	Skills, rules, commands, MCP servers, hooks
GitHub Copilot	✅ Supported	Skills, rules, commands, agents, MCP servers, local hooks
Gemini (CLI/VS Code)	✅ Supported	Skills, rules, commands, MCP servers, hooks
Gemini (JetBrains)	✅ Supported	Rules, MCP servers only (no commands/hooks)
Gemini (Android Studio)	✅ Supported	Rules, MCP-remote only (HTTP, no stdio)
Kiro	✅ Supported	Skills, rules, commands, MCP servers
Openclaw	✅ Supported	Skills, rules, commands
OpenCode	✅ Supported	Skills, commands, agents, rules, MCP servers
claude.ai (web)	✅ Supported	Via the skills.new cloud relay
chatgpt.com (web)	✅ Supported	Via the skills.new cloud relay

Roadmap

• ✅ Local, Git, and Skills.new vaults
• ✅ Claude Code support
• ✅ Cline support
• ✅ Cursor support
• ✅ GitHub Copilot support
• ✅ Gemini support
• ✅ Codex support
• ✅ Kiro support
• ✅ Openclaw support
• ✅ OpenCode support
• ✅ claude.ai and chatgpt.com support via the skills.new cloud relay
• ✅ Org, Team, Bot, Repository & Personal installation targets for all vault types
• ✅ Skill discovery - Use Skills.new to discover relevant skills from your code and architecture
• ✅ Analytics - Track skill usage and impact
• RBAC and change request flow - Support a gated skill update flow

License

See LICENSE file for details.

---

Development

Click to expand development instructions

Documentation

• Vault Spec - Vault directory structure
• Manifest Spec - sx.toml source-of-truth format (assets, scopes, teams)
• Lock Spec - Per-user resolved lock file
• Scoping - Install targets and links to per-scope docs (orgs, repos, teams, users, bots)
• Vault copy - sx vault copy cross-vault migration (assets, teams, bots, scopes, audit, usage)
• Audit log - Event catalog, sx audit filters, storage format
• Usage analytics - sx stats dashboard, JSON output, event format
• Metadata Spec - Asset metadata format
• MCP Spec - MCP server and query tool
• Profiles - Multiple configuration profiles
• Clients - Client support model and IDE vs CLI limitations
• Cloud relay - Expose your vault to claude.ai and chatgpt.com via skills.new
• Library - Use sx as a Go library via the pkg/sxvault public API

Prerequisites

Go 1.25 or later is required. Install using gvm:

# Install gvm
bash < <(curl -s -S -L https://raw.githubusercontent.com/moovweb/gvm/master/binscripts/gvm-installer)

# Install Go (use go1.4 as bootstrap if needed)
gvm install go1.4 -B
gvm use go1.4
export GOROOT_BOOTSTRAP=$GOROOT
gvm install go1.25
gvm use go1.25 --default

Building from Source

make init           # First time setup (install tools, download deps)
make build          # Build binary
make install        # Install to GOPATH/bin

Testing

make test           # Run tests with race detection
make format         # Format code with gofmt
make lint           # Run golangci-lint
make prepush        # Run before pushing (format, lint, test, build)

Releases

Tag and push to trigger automated release via GoReleaser:

git tag v0.1.0
git push origin v0.1.0