Glossary
Pre-execution guard
A check that runs on an AI agent’s proposed tool call before the call executes, and decides whether to allow it, ask a human, or block it.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Pre-execution guard means
A pre-execution guard sits between an agent’s decision to call a tool and the tool actually running. It sees the proposed action (the tool name and its arguments) plus whatever context the caller supplies, such as the user’s request, the recent conversation and any constraints, and returns a decision before anything happens.
The point of checking before execution is that many agent actions cannot be undone: a dropped table, a sent email or an issued refund. Output filters and after-the-fact logging are useful, but they see the damage only once it is done.
MCP Guard is a pre-execution guard. For each check it returns a verdict (allow, ask or block) and nine scores, including P(safe), whether the action is destructive, its blast radius and whether its arguments are grounded in what the user asked. A guard reduces risk; it does not guarantee safety, so it belongs alongside deterministic controls such as least privilege and allowlists.
Related terms
How Pre-execution guard connects to the rest of the vocabulary.