Glossary
Verdict (allow / ask / block)
MCP Guard’s decision for a checked action: allow it, ask a human first, or block it, derived from the model’s scores by a default rule.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Verdict (allow / ask / block) means
Every MCP Guard check returns a verdict with three possible values:
- allow: run the action.
- ask: hold it and ask a human (or another process) to approve.
- block: do not run it.
The default rule on the server is: block if P(unsafe) is at least 0.8 or the approval_policy head says reject; ask if P(unsafe) is at least 0.3, the approval_policy head says require_human, or args_grounded is below 0.5; otherwise allow. The response also includes p_unsafe, short reasons and all nine scores.
The verdict is a starting point, not a fixed policy. Callers can ignore it and apply their own thresholds to the scores, for example asking more often on production systems than in a sandbox. Because scores are only as meaningful as their calibration on your traffic, tuning thresholds on a sample of your own labelled actions is worthwhile.
Related terms
How Verdict (allow / ask / block) connects to the rest of the vocabulary.