OP here. Just to add technical context: I used Upstash Redis because I needed the latency to be negligible (<50ms). The Python decorator wraps the tool function, checks the Redis key for the specific Project ID, and raises a PermissionError if the flag is true.
I'm curious if others are handling this via LLM-based guardrails (checking the input with another LLM) or hard-coded logic like this? In my testing, LLM-guardrails added too much latency.
OP here. Just to add technical context: I used Upstash Redis because I needed the latency to be negligible (<50ms). The Python decorator wraps the tool function, checks the Redis key for the specific Project ID, and raises a PermissionError if the flag is true.
I'm curious if others are handling this via LLM-based guardrails (checking the input with another LLM) or hard-coded logic like this? In my testing, LLM-guardrails added too much latency.