SkillGraph
My SkillsMarketplaceSkill WishesCustom OrdersSafetyGitHubAbout GitHubQ&A
…CR
Create Skill

Skill Safety

What risks SkillGraph checks for, the limits of those checks, Pro certifications, and your responsibility after cloning.

See Q&A

Local execution

SkillGraph does not run AI. All execution after clone happens in your environment.

Threat detection

Pre-screening for injection, malware, secret leaks, and other major risks.

Trust badges

Pre-screened and GitHub-verified badges show trust at a glance.

Agent connection

SkillGraph is a catalog. To use skills, connect your coding agent (Cursor, etc.).

Do I need to link MCP to my LLM agent? Yes. MCP registers SkillGraph tools with your agent (Cursor, Claude Desktop, etc.). SkillGraph does not run AI — your agent fetches Q&A and skill data via MCP and uses it locally.

MCP server connection

Add the SkillGraph MCP server in Cursor Settings → MCP. Your agent can then call tools for Q&A Issues and skill listings.

  • Generate an access key on the Settings page
  • Copy the MCP config JSON into Cursor
  • Ask your agent to check open Q&A Issues
Sign in to connect

My Skills (Clone) connection

Cloning from the marketplace adds skill content (Markdown, GitHub repo, skillgraph.yaml) to My Skills. Your connected agent reads that Markdown to use the skill.

  • Clone imports the Private Layer locally
  • Export Markdown from My Skills
  • GitHub-linked skills include repo URLs
Go to My SkillsBrowse skills

MCP (live fetch) and Clone (local copy) are separate paths. Use both for latest Q&A plus local skill files.

Threats we consider

Prompt injection

Instruction override, persona hijacking, special-token attacks

Malware / harmful scripts

Dangerous shell commands, obfuscated code, remote execution

Secret / API token leaks

API keys, JWTs, or private keys embedded in skill text

Harmful / illegal content

Instructions that facilitate illegal acts or hateful content

Post-clone tampering

Content was safe at publish time but altered locally after clone

Detection methods

✦Secret leaks

Pattern matching (detect-secrets compatible) — Yelp/detect-secrets compatible rules (Pre-screening)

✦Injection

Instruction override / special token detection — SkillGraph Guard / MCP Guard (Pre-screening)

✦Malware

Dangerous command patterns — semgrep p/security-audit compatible (Pre-screening)

✦Hate / discrimination

Blocked phrase patterns — semgrep p/owasp-top-ten compatible (Pre-screening)

✦Harmful / copyright

Semantic moderation — Claude claude-haiku-4-5 (moderation API only) (Pre-screening)

✦GitHub-linked skills

Attestation file via prescribed CI — GitHub Actions + .skillgraph/attestation.json (GitHub verified (planned))

Trust levels

Pre-screened

Passed automated checks at submission. Meets listing standards but does not guarantee safety after clone.

GitHub security verified

Linked repo ran SkillGraph security CI and the attestation file was validated.

バッジの見た目

Pre-screenedGitHubGitHub security verified

Disclaimer and pre-clone agreement

If skill content is tampered with after cloning, if third-party injection is introduced, or if malware runs in your environment, SkillGraph is not liable for resulting damage. We record your agreement to these terms in the pre-clone modal (for legal audit if needed).