Sx: An Open-Source Package Manager for AI Skills, MCPs, and Commands
Sx is an open-source package manager designed for teams using AI coding assistants. It treats AI configurations—custom skills, MCP server configs, slash commands, hooks, agents, and coding standards—as versioned packages that can be centrally managed, scoped, and shared across an organization.
Why Sx?
The problem: top developers create effective AI setups (prompts, MCPs, commands) but those remain siloed on their machines. Current workarounds—copying into each repo (duplication, drift), global configs (context bloat), or manual plugin installs (client lock-in)—don't scale. Sx solves this by treating AI assets as sharable packages with scoped installation.
Quickstart
# macOS/Linux via Homebrew brew tap sleuth-io/tap brew install sxOr via shell script
curl -fsSL https://raw.githubusercontent.com/sleuth-io/sx/main/install.sh | bash
Initialize and add assets
sx init sx add /path/to/my-skill sx install
Scoped Installation
Assets can be installed for specific scopes, avoiding context bloat:
sx install my-skill --org– everyone in the vaultsx install my-skill --repo github.com/acme/infra– only inside that reposx install my-skill --path github.com/acme/infra#docs/– a specific pathsx install my-skill --team platform– team memberssx install my-skill --user [email protected]– a single usersx install my-skill --bot python-backend– a CI/agent identity
Use sx install --dry-run to preview what would be installed.
Cloud Relay
Expose your vault to Claude.ai or ChatGPT via MCP using the skills.new relay:
sx cloud connect sx cloud serve sx cloud status
The relay opens a WebSocket tunnel—vault content stays local.
Analytics & Audit
Track adoption and changes:
sx stats sx stats --since 7d --json sx audit --actor [email protected] --since 30d --event install.set
Works with Existing Claude Code Setups
If you have skills in ~/.claude, sx can version and sync them:
sx add ~/.claude/commands/my-command sx add ~/.claude/skills/my-skill sx add code-review@claude-plugins-official
Sx auto-detects the asset type (skill, command, MCP, agent, hook, rule) and wraps the files with metadata without modifying them.
Sx supports multiple vaults via profiles (sx profile add work, sx profile use work) and works with any AI client that supports MCP or file-based configs—Claude Code, Cursor, GitHub Copilot, Gemini, Kiro, and more.
📖 Read the full source: HN AI Agents
👀 See Also
PhAIL Benchmark Tests VLA Models on Real Warehouse Robot Tasks
PhAIL is a real-robot benchmark that tests four vision-language-action models on bin-to-bin order picking using a Franka FR3 robot. The best model achieved 64 units per hour, compared to 330 UPH for human teleoperation and 1,300+ UPH for human manual work.
KV Cache Reuse for Long Conversations on Apple Silicon Delivers 200x Speedup
A developer implemented session-based KV cache reuse for local LLM inference using Apple's MLX framework, achieving a 200x improvement in time-to-first-token at 100K context length. The approach keeps the KV cache in memory across conversation turns, processing only new tokens.
OpenClaw Integrates Features from Claude Code Leak
An OpenClaw user had their bot analyze the leaked Claude Code (Rust recreation by Instructkr) and selectively ported specific architectural patterns into their OpenClaw setup. The integration focuses on practical improvements like automatic startup continuity, conversation compaction, and a pre-tool/post-tool hook framework.
ATLAS: Open-Source Test-Time Compute Pipeline for Qwen3-14B Achieves Frontier-Level Coding Performance
A college student has developed ATLAS, an open-source test-time compute pipeline built around Qwen3-14B that achieves 74.6% pass@1 on LiveCodeBench v5 problems at ~$0.004 per task in electricity costs. The system is slow for complex problems but offers comparable performance to frontier models like GPT-5 (84.6%) and Claude 4.5 Sonnet (71.4%).