Adapters

Scopra can evaluate policies through model adapters or custom evaluator functions.

OpenAI

Use openai(...) with the official openai SDK client.


import OpenAI from "openai";
import { openai, PolicyPipeline, PromptInjectionPolicy } from "scopra";
 
const pipeline = new PolicyPipeline({
  evaluator: openai(new OpenAI(), "gpt-4.1-mini"),
  policies: [new PromptInjectionPolicy()],
});

Anthropic

Use anthropic(...) with the official @anthropic-ai/sdk client. Anthropic requires max_tokens; Scopra provides a default and lets you override it.


import Anthropic from "@anthropic-ai/sdk";
import { anthropic, PolicyPipeline, SocialEngineeringPolicy } from "scopra";
 
const pipeline = new PolicyPipeline({
  evaluator: anthropic(new Anthropic(), "claude-sonnet-4-5", {
    maxTokens: 2048,
  }),
  policies: [new SocialEngineeringPolicy()],
});

Vercel AI SDK

Use vercel(...) to adapt a Vercel AI SDK language model to Scopra’s SDK-neutral model interface.


import { PolicyPipeline, PromptInjectionPolicy, vercel } from "scopra";
import { openai } from "@ai-sdk/openai";
 
const pipeline = new PolicyPipeline({
  evaluator: vercel(openai("gpt-4.1-mini")),
  policies: [new PromptInjectionPolicy()],
});

TanStack AI

Use tanstack(...) to adapt a TanStack AI text adapter.


import { PolicyPipeline, SocialEngineeringPolicy, tanstack } from "scopra";
 
const pipeline = new PolicyPipeline({
  evaluator: tanstack(adapter),
  policies: [new SocialEngineeringPolicy()],
});

Custom Evaluators

Provide a function when you want full control over evaluation, deterministic tests, or a non-model policy engine.


import type { PolicyEvaluator } from "scopra";
 
const evaluator: PolicyEvaluator = async ({ request, policies }) => {
  return policies.map((policy) => ({
    policyId: policy.id,
    passed: true,
    reason: `Checked ${request.type}.`,
    confidence: 1,
  }));
};

Where To Run Scopra

Run Scopra before sensitive transitions in your agent workflow.

Stage	Use Scopra for
User input	First messages, suspicious follow-ups, support requests, account changes, and commercial pressure.
Model output	Promises, commitments, professional advice, protected content, or data exposure.
Tool calls	Refunds, deletions, writes to external systems, account access, and other side effects.

User-Facing Denial Copy

Use generateViolationResponse(...) when you want model-generated denial copy from a denied decision.