AI agent guardrails
Guardrails for AI agents, in milliseconds
AI agents act on real systems: they send emails, edit records and run commands. A guardrail checks each proposed action before it runs and decides whether to allow it, reject it or ask a human.
Get Early Access to ClassifierHub: 2× credits in your first paid month.
Why a separate decision layer
Asking the same agent to judge its own actions is slow and unreliable. A separate, fast decision model gives an independent verdict with a probability, in a fraction of the time and cost of another LLM call.
The guardrail decision
The agent action guardrail template answers two questions about each proposed action:
- Verdict: allow, review or reject, based on the task and the action.
- Risk: from none to severe, so you can require review above a risk level.
Using it from MCP
With the ClassifierHub MCP server, agents in Claude, Cursor and other MCP clients can call the guardrail as a tool before acting. Deterministic rules such as allow-lists and spending limits should still live in code; the decision layer handles the judgment calls in between.
Frequently asked questions
Related guides
Last updated 2026-09-25. ClassifierHub is an independent product built on top of the Jev decision model, accessed through OpenRouter. It is not affiliated with or endorsed by TypeSafe or OpenRouter.