llms-full.txt@site git:20260222.df87ca1A

Immutable. This exact content never changes and is served at /api/v1/blob/2d9600df59831023.

# SkillSync MCP — Complete Reference

> Security-gated skill management for AI coding assistants. The only MCP server that gates skill installation behind a full security scan.

## What is SkillSync MCP?

SkillSync MCP (`@stranzwersweb2/skillsync-mcp`) is an open-source Model Context Protocol (MCP) server that acts as a security gateway for installing skills (reusable prompt/tool packages) into AI coding assistants like Claude Code. It connects to the SkillsMP marketplace (skillsmp.com) and provides search, security scanning, installation, and cloud sync capabilities.

The key differentiator: SkillSync is the only tool that gates every skill installation behind a comprehensive security scan using 60+ threat detection patterns across 15 categories. Critical threats (prompt injection, reverse shells, credential theft, supply chain attacks) are permanently blocked with no override.

## Quick Facts

- **Package**: @stranzwersweb2/skillsync-mcp
- **Version**: 1.3.0
- **License**: MIT (free and open source)
- **Author**: Aditya Sugandhi (https://adityasugandhi.com)
- **npm**: https://www.npmjs.com/package/@stranzwersweb2/skillsync-mcp
- **GitHub**: https://github.com/adityasugandhi/skillsync-mcp
- **Website**: https://skillsync.js.org/
- **Node.js**: Requires >= 20
- **Transport**: MCP over stdio (StdioServerTransport)
- **Dependencies**: @modelcontextprotocol/sdk, zod

---

## Installation

### One-line install (npx)

```bash
npx -y @stranzwersweb2/skillsync-mcp
```

### Global install

```bash
npm install -g @stranzwersweb2/skillsync-mcp
```

### Claude Code

Add to `~/.claude/settings.json`:

```json
{
  "mcpServers": {
    "skillsmp": {
      "command": "npx",
      "args": ["-y", "@stranzwersweb2/skillsync-mcp"]
    }
  }
}
```

### OpenClaw

Add to `~/.openclaw/mcp.json`:

```json
{
  "mcpServers": {
    "skillsmp": {
      "version": "1.3.0",
      "autoUpdate": false,
      "command": "npx",
      "args": ["-y", "@stranzwersweb2/skillsync-mcp@1.3.0"]
    }
  }
}
```

OpenClaw uses the same SKILL.md format as Claude Code. Pin versions and review tool policies.

### Cursor

Add to `.cursor/mcp.json` in your project root:

```json
{
  "mcpServers": {
    "skillsmp": {
      "command": "npx",
      "args": ["-y", "@stranzwersweb2/skillsync-mcp"]
    }
  }
}
```

### Windsurf

Add to `~/.windsurf/mcp.json`:

```json
{
  "mcpServers": {
    "skillsmp": {
      "command": "npx",
      "args": ["-y", "@stranzwersweb2/skillsync-mcp"]
    }
  }
}
```

### GitHub Copilot

Add to `.github/copilot-mcp.json` (project) or `~/.github/copilot-mcp.json` (global):

```json
{
  "mcpServers": {
    "skillsync": {
      "command": "npx",
      "args": ["-y", "@stranzwersweb2/skillsync-mcp"],
      "env": {
        "SKILLSMP_API_KEY": "your-api-key"
      }
    }
  }
}
```

### Zed

Add to `~/.config/zed/settings.json` under `"context_servers"`:

```json
{
  "context_servers": {
    "skillsync": {
      "command": {
        "path": "npx",
        "args": ["-y", "@stranzwersweb2/skillsync-mcp"],
        "env": {
          "SKILLSMP_API_KEY": "your-api-key"
        }
      }
    }
  }
}
```

### Client Compatibility Table

| Client | Config Path | Skill Format |
|--------|------------|--------------|
| Claude Code | ~/.claude/settings.json | SKILL.md in ~/.claude/skills/ |
| OpenClaw | ~/.openclaw/mcp.json | SKILL.md (same format) |
| Cursor | .cursor/mcp.json | MCP tools only |
| Windsurf | ~/.windsurf/mcp.json | MCP tools only |
| GitHub Copilot | .github/copilot-mcp.json | MCP tools only |
| Zed | ~/.config/zed/settings.json | MCP tools only |
| nanobot | MCP config | MCP tools only |

---

## All 13 MCP Tools (Detailed)

### Search & Discovery

#### 1. skillsmp_search — Keyword Search
Search the SkillsMP marketplace by keyword. Returns skill names, descriptions, authors, star counts, and GitHub links. Supports sorting by stars or recency. All results are sanitized against prompt injection.

Parameters:
- query (string, required): Search keyword(s), 1-200 chars
- limit (number, optional): Max results, 1-100, default 20
- sortBy (enum, optional): "stars" or "recent", default "recent"

#### 2. skillsmp_ai_search — AI Semantic Search
Natural language search powered by Cloudflare AI. Returns relevance-scored results. Finds skills even when keywords do not exactly match.

Parameters:
- query (string, required): Natural language search query, 1-200 chars
- limit (number, optional): Max results, 1-50, default 10

#### 3. skillsmp_search_safe — Search + Auto-Scan
Combines keyword search with automatic security scanning of the top N results. Shows risk levels and threat summaries inline with search results.

Parameters:
- query (string, required): Search keyword(s)
- limit (number, optional): Max search results, default 10
- scanTop (number, optional): How many top results to scan, 1-5, default 3
- sortBy (enum, optional): "stars" or "recent"

#### 4. skillsmp_suggest — AI Suggestions
Recommends new skills based on what you already have installed and optional context about your current project. Filters out already-installed skills.

Parameters:
- context (string, optional): What you're working on, e.g. "React testing"
- limit (number, optional): Max suggestions, 1-20, default 5

### Security

#### 5. skillsmp_scan_skill — Security Scan
Scans a GitHub skill repository for 60+ threat patterns across 15 categories. Returns a risk level (safe/low/medium/high/critical), threat list with line numbers, content hash, files scanned count, and recommendation.

Parameters:
- githubUrl (string, required): GitHub URL to scan

#### 6. skillsmp_compare — Side-by-Side Compare
Compare two skills by GitHub URL or installed skill name. Shows risk levels, threat counts, categories, and a recommendation for which is safer.

Parameters:
- skillA (string, required): GitHub URL or installed skill name
- skillB (string, required): GitHub URL or installed skill name

#### 7. skillsmp_audit_installed — Deep Audit
Forces a fresh security scan on a specific installed skill. Shows detailed threat report, file count, total size, content hash, and SKILL.md status.

Parameters:
- name (string, required): Name of the installed skill

### Skill Management

#### 8. skillsmp_install_skill — Install
Scans a GitHub skill, then installs it to ~/.claude/skills/. Critical threats are permanently blocked. Medium/high risk requires force=true. Runs npm install with --ignore-scripts.

Parameters:
- githubUrl (string, required): GitHub URL
- name (string, optional): Custom skill name
- force (boolean, optional): Override medium/high risk block

#### 9. skillsmp_uninstall_skill — Uninstall
Removes an installed skill directory and updates the in-memory registry.

Parameters:
- name (string, required): Skill name to remove

#### 10. skillsmp_list_installed — List Installed
Lists all installed skills with risk levels, file counts, SKILL.md status, and last scan timestamps.

Parameters:
- refresh (boolean, optional): Force re-sync before listing

### Cloud Sync

#### 11. skillsync_configure — Configure Sync
Manage sync subscriptions and settings. Add/remove search subscriptions, set sync interval (hours), risk threshold (safe/low/medium), and conflict policy (skip/overwrite/unmanage).

Parameters:
- action (enum, required): "add", "remove", "list", or "set"
- query (string, optional): For add action
- subscriptionId (string, optional): For remove action
- syncIntervalHours (number, optional): For set action, 0=manual
- maxRiskLevel (enum, optional): "safe", "low", or "medium"
- conflictPolicy (enum, optional): "skip", "overwrite", or "unmanage"

#### 12. skillsync_sync_now — Run Sync
Polls subscriptions, diffs against installed skills, installs/updates/removes as needed. Supports dry-run mode.

Parameters:
- dryRun (boolean, optional): Preview without changes, default false

#### 13. skillsync_status — Sync Status
Shows sync engine state: managed vs manual skills, subscription count, last sync time, next scheduled sync.

No parameters.

---

## Security Model (Detailed)

### Risk Levels

| Level | Behavior |
|-------|----------|
| Safe / Low | Install proceeds with warnings shown |
| Medium / High | Blocked — requires force=true to override |
| Critical | Permanently blocked — no override possible |

### 15 Threat Categories

**Critical (block installation):**
1. Prompt Injection — override AI instructions, role reassignment, fake system prompts, LLM control tokens, hidden HTML instructions, zero-width Unicode
2. Unicode Homoglyphs — Cyrillic/Greek/IPA/fullwidth characters mixed with ASCII for visual spoofing
3. Destructive Shell — rm -rf /, chmod 777, mkfs, dd, fork bombs
4. Remote Code Execution — curl/wget pipe to shell, secret exfiltration via curl
5. Reverse Shells — bash /dev/tcp, netcat, ncat, socat, Python/PHP/Ruby/Perl one-liners
6. Credential Theft — SSH keys, AWS creds, GPG keys, Kubernetes config, Docker config, .env files, wallet files, keychain
7. Supply Chain — overwrite package.json, write to node_modules, modify CI/CD workflows, npm publish, npm registry redirect, git credential modification
8. Privilege Escalation — sudo, write to /etc/ or /usr/, chown root
9. Exfiltration — secret exfiltration via curl with environment variables

**Warning (flag for review):**
10. Obfuscation — Base64/hex decoding, String.fromCharCode, shell base64 decode pipes
11. Code Execution — eval(), new Function(), dynamic import/require, process.binding, VM module
12. Shell Execution — child_process, execSync/spawnSync, exec with string args
13. Network — fetch to external domains, axios, raw sockets, WebSocket, HTTP servers
14. Crypto Mining — stratum pool connections, xmrig/cryptonight/coinhive/cpuminer
15. Dotfile Poisoning — writing to .bashrc, .zshrc, .npmrc, .yarnrc, .gitconfig

Additional warning categories: DNS exfiltration, crypto wallet addresses, filesystem writes outside project, prototype pollution, time bombs, environment modification, clipboard/input capture.

### Multi-line Pattern Detection

The scanner also applies multi-line patterns to detect evasion attempts where malicious commands are split across lines (e.g., curl pipe to shell with line continuation).

### Safety Guards

- **SSRF Prevention**: Only github.com URLs accepted for scanning
- **Path Traversal Prevention**: Skill names validated against /^[a-zA-Z0-9][a-zA-Z0-9_-]{0,63}$/
- **npm --ignore-scripts**: Blocks postinstall/preinstall script attacks
- **File Limits**: Max 50 files, 512KB per file, 2MB total per skill
- **ReDoS Protection**: Lines over 2000 characters are skipped
- **Content Hashing**: SHA-256 hash for TOCTOU (time-of-check-to-time-of-use) verification
- **Output Sanitization**: Strips zero-width Unicode characters and bidirectional text overrides from all third-party content
- **Scan Cache**: 15-minute TTL with 100-entry LRU eviction
- **Startup Verification**: Background discovery, content hashing, and fs.watch for live filesystem sync
- **Binary Detection**: 30+ binary extensions skipped automatically
- **Suspicious Filenames**: postinstall.sh, preinstall.js, .env, .npmrc, makefile flagged

---

## Architecture

### Data Flow

```
MCP Client (stdio) → tools.ts (13 tool handlers) → Business Logic → Response
                                                     ├── api-client.ts (SkillsMP REST API)
                                                     ├── security-scanner.ts (GitHub fetch + threat scan)
                                                     ├── skill-manager.ts (installed skill registry + fs.watch)
                                                     ├── installer.ts (filesystem + npm install)
                                                     ├── sync-engine.ts (cloud sync orchestration)
                                                     └── sanitize.ts (output safety)
```

### Startup Sequence

1. McpServer created
2. registerTools(server) — registers all 13 tools
3. server.connect(transport) — MCP is live, tools available immediately
4. skillManager.initialize() — BACKGROUND (fire-and-forget)
   - Discover skills in ~/.claude/skills/
   - Security scan each skill
   - Start fs.watch for live additions/removals

### Key Files

| File | Purpose |
|------|---------|
| src/index.ts | Entry point — creates McpServer, connects stdio |
| src/tools.ts | 13 MCP tool definitions with Zod schemas |
| src/skill-manager.ts | Discovery, local scanning, registry, fs.watch |
| src/api-client.ts | SkillsMP marketplace REST API client |
| src/security-scanner.ts | Core threat engine — GitHub fetch + pattern matching |
| src/installer.ts | Skill install/uninstall — path validation, file writing |
| src/patterns.ts | 60+ regex threat patterns (CRITICAL + WARNING) |
| src/constants.ts | Config: skills dir, file limits, extensions |
| src/sanitize.ts | Anti prompt injection output cleaning |
| src/sync-engine.ts | Cloud sync orchestration |
| src/sync-config.ts | Sync subscription management |
| src/sync-lock.ts | Managed skill lock file |

---

## Why SkillSync vs Alternatives

| Capability | Raw git clone | Other Tools | SkillSync MCP |
|-----------|--------------|-------------|---------------|
| Security scan before install | No | No | Yes — 60+ patterns |
| Blocks critical threats | No | No | Yes — prompt injection, RCE, credential theft |
| Multi-client support | N/A | Varies | 7 clients |
| Marketplace search | Manual | Some | Keyword + AI semantic |
| Cloud sync | No | No | Subscription-based sync |
| Startup verification | No | No | fs.watch + content hash |
| Output sanitization | No | No | Anti prompt injection |

---

## Development

```bash
git clone https://github.com/adityasugandhi/skillsync-mcp.git
cd skillsync-mcp
npm install
npm run build
npm run dev       # Watch mode
npm run test:build  # Build + run tests (89 test cases)
```

## Contributing

See CONTRIBUTING.md. Browse open issues or look for the good-first-issue label. All PRs must pass the existing test suite.