| name | mcp-tool-policy-review |
| description | Use when adding, changing, or reviewing MCP servers, agent tools, provider integrations, credentials, scopes, or tool approval behavior. |
MCP Tool Policy Review
Workflow
- Read
AGENTS.md, docs/MCP_TRUST_BOUNDARY.md, and the affected tool policy.
- Classify each tool as read-only, external write, destructive, credential-bearing, or expensive.
- Verify the policy names allowed operations, denied operations, scopes, approval behavior, logging, and validation.
- Check that secrets are not stored in prompts, docs, screenshots, or example configs.
- Confirm permission enforcement lives in code or configuration, not only prompt text.
- Recommend tests or manual validation for the trust boundary.
Required Findings
Report a finding when:
- destructive or external-write tools lack approval gates
- tool scopes are vague or impossible to enforce
- credentials can leak into model context or logs
- examples contain real secrets or private endpoints
- logging omits necessary audit context or stores sensitive payloads
Output
Lead with findings. Include the affected tool, risk, file reference, and concrete fix.