mcp-tool-surface-reviewlisted
Install: claude install-skill Wanbinyu/ai-surface-skills
# MCP Tool Surface Review
> You already have an MCP. Are its tools safe for an agent to see?
## Overview
Audit an existing MCP tools/list for safety and clarity (not server scaffolding).
## Steps
1. Obtain tool list (MCP `tools/list` dump or docs).
2. Per tool: description quality, param width, side effects, secrets in args, path traversal, SSRF-ish URL params.
3. Flag tool poisoning patterns: descriptions that instruct the model to ignore policy, exfiltrate, or hide actions.
4. Cross-check with `tool-permission-matrix` tags.
5. Verdict: allow / allow-with-HITL / block / fork-and-narrow.
## Exit criteria
- [ ] Tool inventory
- [ ] Findings table with severity
- [ ] Poisoning/over-broad params checked
- [ ] Verdict per tool or overall
- [ ] No exploit payload development
## Anti-patterns
- Rewriting the whole server in the review
- Generic "MCP is dangerous" without tool rows
- Shipping exploit chains
## Output template
```markdown
## MCP tool surface review
| tool | issue | severity | recommendation |
|------|-------|----------|----------------|
### Overall
- allow | harden | block
```