MCP Guard blocks dangerous AI agent actions using a simple YAML config file. A stdio JSON-RPC proxy that intercepts tools/call, classifies decisions (allow/block/hold), and enforces permission policies with human-in-the-loop approval.
mcp-guard exposes 8 tools with mixed quality. Tools 1-6 (actions-mcp and infra-mcp) have well-structured schemas with clear descriptions and _pp metadata support. Tools 7-8 (example/server.ts) are severely underdeveloped: delete_user_data and deploy_production lack parameter descriptions, have minimal input schemas, and no error handling guidance. The server is a permission-gating proxy, not a primary tool provider, which limits its definition scope. Naming is verb-first and clear (send_email, post_x, spend, merge_pr, run_sql_migration, deploy). However, tools 7-8 violate baseline standards: descriptions under 20 chars, missing parameter docs, no output schema documentation.
Delete all user data
Deploy a service to an environment (reversible — a rollback exists; guarded as a Decide upstream).
Deploy to production
Merge a GitHub pull request (reversible — a merge can be reverted; guarded as a Decide with optional countdown upstream). Carries scope_ref/scope_sha so the upstream signer can bind a receipt to this exact merge.
Post to X / Twitter (public, irreversible — hard-blocked until a signed receipt is issued upstream).
Run a SQL schema migration (irreversible — schema change is a hard stop until approved upstream).
Send an outbound email (outbound comms; reversible — guarded with a countdown upstream).
Tools 7-8 (delete_user_data, deploy_production) have descriptions under 20 characters ('Delete all user data', 'Deploy to production'), violating baseline standards.
delete_user_data and deploy_production lack parameter descriptions. 'user_id' and 'version' have no explanation of format, constraints, or purpose. LLMs cannot infer whether to pass UUIDs, emails, or opaque IDs.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Tools 1-6 return text results (e.g. 'Email sent to...'), but structure is not formally declared.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Spend money / make a payment (reversible below cap; verified/decided upstream by amount).
Destructive/irreversible tools (post_x, run_sql_migration, delete_user_data) lack confirmation or dry-run patterns. Agents can trigger irreversible actions without explicit user approval.
Error handling is minimal. Tool stubs in actions-mcp/server.mjs return generic success messages. No guidance on retryability, user-fixable errors, or recovery steps. LLMs cannot self-correct on failures.