One trust check before any install.

Install the gate

Drop one file into your agent's skills directory. For Claude Code (user-scope skills):

mkdir -p ~/.claude/skills/pre-install-trust-gate && \
curl -fsSL https://mcpskills.io/skills/pre-install-trust-gate/SKILL.md \
  -o ~/.claude/skills/pre-install-trust-gate/SKILL.md

For OpenClaw, Hermes, Cursor, or any agent that loads SKILL.md files, fetch the same file into that agent's skills folder. The skill is plain Markdown with standard frontmatter — no runtime, no dependency to install.

Canonical source: /skills/pre-install-trust-gate/SKILL.md (CC BY 4.0).

What it does

Before the agent installs, adds, or enables anything third-party, it makes one request — GET /api/gate/<target> — and acts on the boolean:

curl -s "https://mcpskills.io/api/gate/modelcontextprotocol/servers"
# → { "proceed": true, "reason": "Established: composite 6.8, no disqualifiers", "tier": "established", ... }

curl -s "https://mcpskills.io/api/gate/npm:left-pad"          # npm package
curl -s -H "X-API-Key: $KEY" "https://mcpskills.io/api/gate/owner/repo"   # higher daily limit

Verdict	What the agent does
proceed: true	Install proceeds. The agent notes the tier and score in one line.
proceed: false	Stops. Surfaces the reason and the score page, asks for an explicit "install anyway."
429 / error	Trust check couldn't complete — fails to the human, never silently installs un-checked.

The gate returns proceed: true for Verified and Established with no hard disqualifier, and proceed: false for anything Blocked or New — including no license, a critical/KEV CVE, a known-malicious advisory, token exfiltration in CI, or unsafe patterns in source. Same verdict the auto_gate MCP tool returns; the gate is just that primitive on a plain URL.

Try it

Live against the same endpoint your agent calls. Free tier is 10 checks/day per IP — an API key raises it.

It passes its own scan

The trust gate is the rare skill that practices what it checks. Run it through the same engine and it earns the full 2.0 / 2.0 security-transparency bonus (it declares its network, permissions, and credentials in frontmatter) and a clean tool-safety pass — zero injection, exfiltration, or shell findings.

Why a gate

Most install flows run unreviewed third-party code with full local permissions. The published record makes the case:

• ClawHavoc shipped 1,184 malicious AI skills — 7,700 installs before takedown.
• CVE-2025-6514 gave any malicious MCP server RCE on the client through mcp-remote (CVSS 9.6), across 437,000+ downloads before the patch.
• When OX Security submitted a benign proof-of-concept malicious MCP server to 11 registries, 9 accepted it with no security review.

A five-second pre-install gate is the cheapest control that catches the obvious cases. It is not a guarantee — it is static, pre-install evidence. Pair it with the controls a scanner can't replace: sandboxing, manifest hashing, version pinning. See how trust scoring works and the State of Hermes Skill Security report.