verdicter 1.0.0

Runtime security for AI agents — evaluate tool calls against policies in real time

verdicter

Runtime security for AI agents. Evaluate every tool call against your policies — get ALLOW, DENY, MODIFY, or ESCALATE in under 50ms.

→ Get your free API key at verdicter.dev

---

Install

pip install verdicter

With LangChain support:

pip install "verdicter[langchain]"

---

Quick start

import asyncio
from verdicter import Verdicter

client = Verdicter(api_key="verdicter_live_...")

async def main():
    result = await client.evaluate(
        agent_id="support_bot",   # registered in your Verdicter dashboard
        tool="send_email",
        payload={"to": user_email, "subject": subject, "body": body},
    )

    if result.decision == "ALLOW":
        await send_email(payload)
    elif result.decision == "MODIFY":
        await send_email(result.modified_payload)   # Verdicter rewrote the payload
    elif result.decision == "DENY":
        raise PermissionError(f"Blocked: {result.reason}")
    elif result.decision == "ESCALATE":
        await request_human_approval(payload)

---

Wrap a function

Zero per-call boilerplate — evaluation happens automatically on every invocation:

from verdicter import Verdicter, WrapOptions

client = Verdicter(api_key="verdicter_live_...")

safe_send_email = client.wrap(
    "send_email",
    send_email,
    WrapOptions(agent_id="support_bot"),
)

# Policy enforcement is automatic
await safe_send_email(to=user_email, subject=subject, body=body)

---

LangChain adapter

from verdicter import Verdicter
from verdicter.adapters.langchain import wrap_tools
from langchain_core.tools import tool

client = Verdicter(api_key="verdicter_live_...")

@tool
def send_email(to: str, subject: str, body: str) -> str:
    """Send an email."""
    ...

# Wrap your tools — every invocation goes through Verdicter
safe_tools = wrap_tools([send_email], client, agent_id="support_bot")

agent = create_react_agent(llm=llm, tools=safe_tools)

---

Configuration

client = Verdicter(
    api_key="verdicter_live_...",
    timeout=5.0,        # seconds, default 5.0
    max_retries=2,      # default 2
    fail_open=False,    # if True, ALLOW on network errors (default: False = fail closed)
)

---

Decisions

Decision	Meaning
ALLOW	Policy passed — run the tool
DENY	Policy blocked it — don't run
MODIFY	Policy rewrote the payload — use result.modified_payload
ESCALATE	Needs human review — route to your approval flow

---

Links

• Sign up free → verdicter.dev
• Dashboard → app.verdicter.dev
• Documentation → docs.verdicter.dev
• npm package → npmjs.com/package/verdicter

---

MIT License © Verdicter