SDK Wrappers

agents-chain ships Proxy wrappers for the OpenAI and Anthropic SDKs. These intercept API calls and gate them through the auth pipeline.

OpenAI Wrapper

import OpenAI from 'openai';
import { AppChain } from 'agents-chain';

const chain = await AppChain.create({
  providerName: 'openai-service',
  issuer: 'https://myservice.com',
  capabilities: [
    {
      name: 'chat.completions.create',
      description: 'Create a chat completion',
      inputSchema: { type: 'object', properties: {} },
      outputSchema: { type: 'object' },
    },
  ],
});

const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
const grants = [
  { capability: 'chat.completions.create', status: 'active' },
];

// Wrap the SDK — method paths are mapped to capability names
const secured = chain.openai(openai, grants);

// This goes through the auth pipeline before reaching OpenAI
const response = await secured.chat.completions.create({
  model: 'gpt-4',
  messages: [{ role: 'user', content: 'Hello' }],
});

Anthropic Wrapper

import Anthropic from '@anthropic-ai/sdk';
import { AppChain } from 'agents-chain';

const chain = await AppChain.create({
  providerName: 'anthropic-service',
  issuer: 'https://myservice.com',
  capabilities: [
    {
      name: 'messages.create',
      description: 'Create a message',
      inputSchema: { type: 'object', properties: {} },
      outputSchema: { type: 'object' },
    },
  ],
});

const anthropic = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
const grants = [
  { capability: 'messages.create', status: 'active' },
];

const secured = chain.anthropic(anthropic, grants);

const response = await secured.messages.create({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello' }],
});

How It Works

The wrappers use JavaScript Proxy to intercept nested property access. When you call secured.chat.completions.create(args), the wrapper:

1. Maps the path chat.completions.create to a capability name
2. Runs the full auth pipeline (token mint → 11-step verify → constraints)
3. Calls the real SDK method if auth passes
4. Records the call in the audit log

Constraint enforcement works on the SDK call arguments. You can restrict which models, parameters, or inputs an agent is allowed to use.