OpenRouter Armor OpenRouter Armor is a beta plugin that guards AI agent tool calls before they execute. It scores each proposed action against user-defined natural-language rules plus conversation context using TypeSafe’s Jev decision model, blocking out-of-policy calls more accurately and cheaply than regex filters or Claude Code’s auto classifier. Any agent routing inference through OpenRouter can enable it; setup with Claude Code is a one-line command.

Visit product

Featured media

Cosigns Received

B_a_aB
“Openrouter got some really nice addition, really nice feature, it is more exciting that they based their solution using classifier approach - of course it is Jev. #TypeSafeAi #jev #openrouter”
Source
Truong Pham
“Checking tool calls before they run is the right layer. Most agent damage isn't a bad answer, it's one delete or send call nobody looked at.”
Source
Vlad Gerasimchuk
“Guarding tool calls before execution is the right layer for agents that touch real systems. Regex filtering misses intent, so scoring each call against rules plus transcript context makes sense. As a solo builder running agents on my own projects, this is the kind of safety net I want by default.”
Source
Tom Refactored
“Scared of prompt injection? Turn this on Awesome job OpenRouter”
Source
Xiaodao AI
“Guarding tool calls before they hit the real world is the missing layer. Browser/checkout agents especially — policy at the gateway beats hoping the model “just behaves.””
Source
Solomon
“Separating hard_deny (immune to user clearance) from soft_deny is the crucial pattern here. Otherwise an indirect prompt injection in a fetched file can just gaslight the evaluator into thinking the user authorized the destructive call. Putting a 280ms specialized model between plan and dispatch is way saner than praying regex catches sneaky bash flags :)”
Source
HiloW.Ai
“Regex on tool calls was always theater. The guard has to read intent, not strings. A cheap dedicated model between plan and execute is the layer most agent stacks skip. The real test: chains where every single call looks harmless on its own.”
Source