SG SkillGuard

SkillGuard

Scan skills and MCP servers locally. Stop changed or unscanned tools at runtime.

Read docs
Local install no cloud engine required
$ curl -fsSL https://skillguard.sh/install | sh
$ skillguard --target all --engines local
runtime gate PreToolUse hook
SkillGuard terminal demo showing high-risk findings before runtime
A real terminal demo from this repository. The hook refuses unsafe skill use before the tool runs.

Agent tooling is a supply chain now.

Skills and MCP servers can read files, launch processes, and inject instructions. SkillGuard treats them like executable dependencies.

341

malicious skills found in one marketplace audit

Koi Security reported typosquatted ClawHub skills that installed an infostealer and targeted wallets, SSH keys, and env files. Read the report

13.4%

contained critical issues

Snyk found critical-level security issues in a follow-up study of skills on the same registry. Read the study

Scan-only tools leave a gap

A scanner helps when you remember to run it. SkillGuard adds a runtime gate, so changed or unscanned targets stop at invocation.

Secrets stay redacted

MCP env and header values are inspected in memory. Cache records keep names and metadata, not secret values.

Local first by default

The local engine needs no token. Optional engines such as Cisco, Snyk, and SkillAudit can be added when a deeper pass is worth it.

The gate has three jobs.

SkillGuard reads the tool surface, decides whether it cleared, then records the decision for later review.

Read the whole tool surface

Skill files, bundled scripts, JSON and YAML config, MCP launch commands, endpoint metadata, env names, and header names are scanned together.

Refuse what has not cleared

Claude Code hook checks happen before tool use. MCP enforcement can warn or block unknown, changed, high-risk, and critical verdicts.

Remember every decision

Local history tracks scans, hook allows, blocks, acknowledgements, and false-positive marks in a SQLite database.

Install once. Gate every run.

Start with the local installer, or choose the agent surface you already use.

Copies files, patches supported settings, and keeps a backup before writing.

Local
$ curl -fsSL https://skillguard.sh/install | sh
$ skillguard --target all --engines local

Questions security teams ask first.

What does SkillGuard scan?
Skills, bundled files, MCP server config, launch commands, endpoint metadata, env names, and header names.
Does it block MCP tool calls?
Yes. The hook can enforce MCP policy and refuse unknown, unscanned, high-risk, or critical server verdicts.
Do secret values touch disk?
No. Env and header values are inspected in memory. Persistent records keep redacted material only.
What leaves my machine?
Nothing with the local engine. Files leave only when you enable optional cloud or remote engines.
Can findings be triaged?
Yes. Use skillguard --review for interactive triage or mark reviewed false positives from the CLI.