GitHub - sleuth-io/sx: sx is a package manager for AI coding assistants

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:

• Share across projects and teams — manage an asset once, install it into any repos or teams
• Share across clients, including the web ones — one asset installs into every major AI assistant
• No Git knowledge required — gives non-technical users easy access; the plumbing stays hidden
• Install the right assets, not all of them — scope to an org, repo, path, team, bot, or user, no context bloat
• Version, observe, and govern — update once; track adoption with sx stats; audit access with sx audit

App or command line — your choice

sx has two front doors to the same libraries: a desktop app and a command line. They share one configuration, so a library added in one shows up in the other — engineers can live in the CLI while everyone else uses the app, and nothing drifts.

If you're not a command-line person, start with the app. It covers the whole workflow without a terminal: create and edit skills in a built-in editor (or just drag files in), publish them to your team's library, install them into your AI tools with one click, and organize everything with collections and teams. Technical users can turn on per-library repository views to see exactly which repos each skill is scoped to.

First launch asks one question — how will you use it? — and sets the library up to match:

Download the app for macOS, Windows, or Linux from Releases (the sx-app-* artifacts). Power users: keep reading for the CLI.

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:

From here the workflow is three steps — manage your assets, distribute them, and govern what ships.

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 alice@acme.com               # 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 alice@acme.com --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

Distribution models

Choose the right distribution model for your team:

Local (Personal)

Perfect for easily sharing personal tools across multiple personal projects

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

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

Git vault (Small teams)

Share assets through a shared git vault

sx init --type git --repo-url git@github.com:yourteam/skills.git

Skills.new (Large teams and enterprise)

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

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 git@github.com: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

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

# 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

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

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)

git tag v0.1.0
git push origin v0.1.0