skillsmp-mcp 0.1.0

MCP server for SkillsMP: search, read-only GitHub fetch, full Cisco AI Skill Scanner, and scan-gated install.

skillsmp-mcp

A Python MCP server for SkillsMP that you own end to end: search the catalogue, read a skill's SKILL.md straight from GitHub, run the full Cisco AI Skill Scanner (including the LLM semantic layer that "lite" wrappers skip), and install skills only after they pass the scan.

Why this exists

SkillsMP publishes a REST API, not an official MCP server. Every skillsmp-mcp-* package on npm is third-party code. This server talks to SkillsMP with your key over exactly one host, fetches skills read-only from GitHub, and gates install on a real security scan — so there's nothing to trust but code you can read.

Tools

Tool	What it does
search_skills(query, limit, sort_by)	Keyword search (sort_by: stars/recent).
read_skill(repo, skill_name)	Fetch SKILL.md from a GitHub owner/repo, read-only. Never scans, never spawns a subprocess, never installs.
scan_skill(repo, skill_name)	Full Cisco scan; reports findings only.
install_skill(repo, skill_name, force=False)	Scan-gated install. Refuses on HIGH/CRITICAL unless force=True.

Requirements

• Python 3.10+
• A SkillsMP API key (sk_live_...) from skillsmp.com — the only hard requirement.
• For scan_skill / install_skill: either uv installed (the server runs the scanner via uvx on demand) or the cisco-ai-skill-scanner package installed directly. Without either, scans return SKIPPED and installs are blocked.
• Recommended: an Anthropic API key (sk-ant-...) to enable the LLM semantic analyzer.

Install

git clone https://github.com/Asyboi/skillsmp-mcp.git
cd skillsmp-mcp

# create an isolated env and install (uv shown; plain venv + pip works too)
uv venv
uv pip install -e .

# optional: install the scanner directly instead of relying on uvx
uv pip install cisco-ai-skill-scanner

This installs two ways to launch the server (use whichever your client prefers):

• the console script skillsmp-mcp (on PATH inside the venv), or
• python -m skillsmp_mcp — portable, no PATH juggling.

Quick smoke test (no keys needed — lists the 4 tools over stdio, then exits):

printf '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"t","version":"0"}}}\n' \
  | SKILLSMP_API_KEY=dummy .venv/bin/skillsmp-mcp

Configure & register with Claude Desktop

Keys are supplied through the MCP server's env block in your Claude config — the code reads them from the process environment, which the client injects at launch. There is no .env file to manage.

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json). Merge into mcpServers — don't replace the block.

"skillsmp": {
  "command": "/absolute/path/to/skillsmp-mcp/.venv/bin/skillsmp-mcp",
  "env": {
    "SKILLSMP_API_KEY": "sk_live_...",
    "SKILL_SCANNER_LLM_API_KEY": "sk-ant-...",
    "SKILL_SCANNER_LLM_MODEL": "anthropic/claude-sonnet-5"
  }
}

SKILLSMP_API_KEY is the only hard requirement. Set SKILL_SCANNER_LLM_API_KEY too so the semantic analyzer runs — without it, scans fall back to static + behavioral only and can miss prose-based prompt injection. All other variables (see the table below) are optional and go in the same env block.

Use an absolute path for command. GUI apps like Claude Desktop don't inherit your shell PATH, so a bare skillsmp-mcp often won't resolve. Point at the venv binary (shown above) — or use "command": "/abs/path/.venv/bin/python", "args": ["-m", "skillsmp_mcp"].

For the same reason, if you rely on uvx to run the scanner, set SKILLSMP_SCANNER_CMD to its absolute path so scanning works under the client:

"SKILLSMP_SCANNER_CMD": "/Users/you/.local/bin/uvx --from cisco-ai-skill-scanner skill-scanner"

Fully quit and relaunch Claude Desktop after editing.

Environment variables

Var	Required	Default	Purpose
SKILLSMP_API_KEY	yes	—	SkillsMP REST auth (sent only to skillsmp.com).
SKILL_SCANNER_LLM_API_KEY	recommended	—	Enables the LLM semantic analyzer.
SKILL_SCANNER_LLM_MODEL	no	anthropic/claude-sonnet-5	LiteLLM model string.
GITHUB_TOKEN	no	—	Read-only PAT; raises GitHub rate limit.
SKILLSMP_INSTALL_DIR	no	~/.claude/skills	Install target.
SKILLSMP_BLOCK_SEVERITIES	no	HIGH,CRITICAL	Severities that block install.
SKILLSMP_SCANNER_CMD	no	auto-detect	Override the scanner command.
SKILLSMP_SCANNER_POLICY	no	—	Cisco policy preset (e.g. strict).
SKILLSMP_SCANNER_TIMEOUT	no	300	Scan timeout (seconds).
SKILLSMP_MAX_FILES / SKILLSMP_MAX_SINGLE_FILE_BYTES / SKILLSMP_MAX_TOTAL_BYTES	no	100 / 512000 / 5242880	Fetch/scan size caps.
VIRUSTOTAL_API_KEY / AI_DEFENSE_API_KEY	no	—	Optional cloud engines.

For standalone runs outside a Claude client (e.g. terminal testing), just export the variables in your shell first — the server reads the same environment either way.

Security model

• Your SkillsMP key is sent only to skillsmp.com. GitHub fetches are unauthenticated (or use a read-only GITHUB_TOKEN); the two never mix.
• Skill files are fetched read-only, written to a private temp dir for scanning, and that dir is deleted afterward. Cisco's analysis is static/AST — nothing is executed.
• install_skill refuses HIGH/CRITICAL findings unless forced, refuses to install unscanned (scanner missing) unless forced, and never silently overwrites — re-installing an existing skill folder also requires force=True.
• Installs are namespaced as <owner>-<repo>__<skill-dir> so skills from different sources can't collide and a crafted name can't escape the install root.
• A clean scan is not a safety guarantee — it means no known patterns matched. Read the SKILL.md yourself; read_skill labels it untrusted for that reason.

Where to add your own touches

• config.py — new env tunables, stricter BLOCK_SEVERITIES, install target.
• scanner.py — analyzer selection, extra engines, custom result parsing.
• installer.py — post-install hooks (e.g. write an index entry to a vault).
• server.py — register more @mcp.tool() functions.

Development

uv pip install -e ".[dev]"
pytest                       # full suite (unit + parsing + gating)
python -m py_compile src/skillsmp_mcp/*.py

Tests are offline by default — network clients are driven with httpx.MockTransport and the scanner logic is unit-tested with fake reports, so pytest needs no keys or the Cisco scanner.

License

MIT — see LICENSE.