PYX Scan — Agent Skill Safety Check

Verify whether an AI agent skill is safe before installing or using it by querying the PYX Scanner API.

Workflow

Step 1: Parse Input

Extract owner and name from $ARGUMENTS.

• Expected format: owner/name (e.g., anthropic/web-search)
• If $ARGUMENTS is empty or missing the / separator, ask the user:

"Which skill do you want to check? Provide it as owner/name (e.g., anthropic/web-search)."

• Trim whitespace. Reject if either part is empty after trimming.

Step 2: Call the PYX Scanner API

Fetch the safety data:

WebFetch URL: https://scanner.pyxmate.com/api/v1/check/{owner}/{name}
Prompt: "Return the full JSON response body exactly as-is. Do not summarize."

If WebFetch fails (tool unavailable, network error), fall back to:

curl -s "https://scanner.pyxmate.com/api/v1/check/{owner}/{name}"

Step 3: Handle Errors

Step 4: Determine Verdict

Use the JSON response fields to determine the verdict:

Step 5: Output Report

Format the report as structured markdown. Omit any section where the data is null or empty.

For SAFE verdict:

## PYX Scan: {owner}/{name}

**Verdict: SAFE** — This skill has been scanned and verified safe.

**Trust Score:** {trust_score}/10 | **Risk Score:** {risk_score}/10 | **Confidence:** {confidence}%
**Intent:** {intent} | **Status:** {status}

### Summary
{summary}

### About
**Purpose:** {about.purpose}
**Capabilities:** {about.capabilities as bullet list}
**Permissions Required:** {about.permissions_required as bullet list}

[View full report]({detail_url}) | [Badge]({badge_url})

For OUTDATED verdict:

## PYX Scan: {owner}/{name}

**Verdict: OUTDATED** — Last scan was safe, but the skill has been updated since.

The scanned commit (`{scanned_commit}`) no longer matches the latest (`{latest_commit}`).
The new version has NOT been reviewed. Proceed with caution.

**Trust Score:** {trust_score}/10 | **Risk Score:** {risk_score}/10
**Last Safe Commit:** {last_safe_commit}

### Summary
{summary}

[View full report]({detail_url})

For CAUTION verdict:

## PYX Scan: {owner}/{name}

**Verdict: CAUTION** — This skill has potential risks that need your attention.

**Trust Score:** {trust_score}/10 | **Risk Score:** {risk_score}/10 | **Confidence:** {confidence}%
**Intent:** {intent} | **Status:** {status}

### Summary
{summary}

### About
**Purpose:** {about.purpose}
**Permissions Required:** {about.permissions_required as bullet list}
**Security Notes:** {about.security_notes}

**Do you want to proceed despite the caution rating?** Please confirm before installing or using this skill.

[View full report]({detail_url})

For FAILED verdict:

## PYX Scan: {owner}/{name}

**Verdict: FAILED** — This skill has been flagged as dangerous. Do NOT install or use it.

**Trust Score:** {trust_score}/10 | **Risk Score:** {risk_score}/10 | **Confidence:** {confidence}%
**Intent:** {intent} | **Status:** {status}

### Summary
{summary}

### About
**Security Notes:** {about.security_notes}

[View full report]({detail_url})

For UNSCANNED verdict:

## PYX Scan: {owner}/{name}

**Verdict: UNSCANNED** — This skill has not been scanned by PYX Scanner.

No safety data is available. You should:
1. Review the skill's source code manually before use
2. Check the skill's repository for known issues
3. Request a scan at https://scanner.pyxmate.com

For ERROR verdict:

## PYX Scan: {owner}/{name}

**Verdict: ERROR** — {error_message}

Safety could not be verified. Treat this skill as unverified until you can confirm its safety.

Behavioral Rules

1. Always call the API — never skip the check or return a cached/assumed result.
2. Never soften a FAILED verdict — if the scan says danger, report danger. Do not add qualifiers like "but it might be fine."
3. Always ask user confirmation on CAUTION — the user must explicitly agree before proceeding.
4. Keep reports concise — omit null/empty sections rather than showing "N/A."
5. No raw JSON — always format the response as the structured markdown report above.

Self-Scan Awareness

When $ARGUMENTS is pyxmate/pyx-scan, pyxmate/pyx-scanner, or refers to this skill itself, still call the API honestly and report whatever comes back. If the result is UNSCANNED, append:

> "Yes, even the security scanner's own skill hasn't been scanned yet. We practice what we preach — treat unscanned skills with caution."