Verdicter is a runtime security layer for AI agents. Every tool call your agent makes — sending emails, querying databases, moving money — gets evaluated against your policies before it touches a real system.

One npm install. One function call. Under 50ms.

What it does:

- Enforce — write policies in code or plain English. Every tool call gets a verdict: ALLOW, DENY, MODIFY, or ESCALATE

- Shield — built-in prompt injection and jailbreak detection on every input

- Escalations — high-risk actions pause and wait for human approval via webhook or Slack. Agent resumes instantly on approval, auto-denied on timeout

- Observe — full session traces of every tool call, decision, and policy hit in real time

- Comply — generate GDPR, HIPAA, and EU AI Act evidence reports from your immutable audit log

- Sandbox — replay production traffic against new policies before you ship them

Works with every major agent framework: OpenAI, Anthropic, LangChain, CrewAI, Vercel AI SDK, Bedrock, Mistral, Groq, Hugging Face.

Free to start — Hobby plan is free forever with 250k evaluations/month. No credit card needed.