Use Vercel AI SDK with AnonRouter

Use @anonrouter/ai-sdk-provider from a server-side JavaScript or TypeScript application. It uses the private ticket flow by default. The inference key stays on the control origin, and each inference request carries a single-use ticket.

Install and configure

These examples require Vercel AI SDK 7 and Node.js 22 or newer.

npm install ai@7 @anonrouter/ai-sdk-provider

Create an inference key in your dashboard and set ANONROUTER_API_KEY in your server environment. Keep the key on your server. Choose a model from the model catalog. For the tool example, choose a route that supports tool calling.

Create provider.ts in your server code.

import { anonrouter } from '@anonrouter/ai-sdk-provider';

export const model = anonrouter('deepseek/deepseek-v3.2');

Set maxRetries: 0 on AI SDK calls. A disconnect can happen after a request has been charged, so repeating it can create another charge.

Stream an answer

Run this in your server application, using provider.ts above:

import { streamText } from 'ai';
import { model } from './provider';

const result = streamText({
  model,
  prompt: 'Describe a quiet morning in two sentences.',
  maxOutputTokens: 256,
  maxRetries: 0,
  onError: () => {
    process.stderr.write('The model stream failed.\n');
    process.exitCode = 1;
  },
});

for await (const text of result.textStream) {
  process.stdout.write(text);
}

This prints the answer for a terminal demonstration. In a web app, send it to the user instead of recording it in server logs. Handle stream errors separately; partial text is not proof that a request completed successfully. See the streaming reference.

Run a bounded tool loop

Your application executes tools. AnonRouter forwards the model's tool calls and the results you send back. Install zod@4 for this unit conversion example.

import { generateText, isStepCount, tool } from 'ai';
import { z } from 'zod';
import { model } from './provider';

const result = await generateText({
  model,
  prompt: 'Use the conversion tool to convert 21 Celsius to Fahrenheit.',
  maxOutputTokens: 256,
  maxRetries: 0,
  stopWhen: isStepCount(2),
  tools: {
    celsiusToFahrenheit: tool({
      description: 'Convert a temperature from Celsius to Fahrenheit.',
      inputSchema: z.object({ celsius: z.number().min(-273.15).max(10000) }),
      execute: async ({ celsius }) => ({ fahrenheit: celsius * 9 / 5 + 32 }),
    }),
  },
});

if (!result.text) throw new Error('No final answer within the two-step limit.');
process.stdout.write(result.text);

The loop allows at most two model steps, including the follow-up after a tool result. Each model request uses your account's available funding. Keep tools narrowly scoped and validate inputs before accessing files or services. See the tool-calling reference.

Create an embedding

import { anonrouter } from '@anonrouter/ai-sdk-provider';
import { embed } from 'ai';

const { embedding } = await embed({
  model: anonrouter.embedding('qwen/qwen3-embedding-0.6b'),
  value: 'The Pacific Ocean is the largest ocean on Earth.',
  maxRetries: 0,
});

Choose how a request is routed

Set a policy when you select a model, or for a single call under providerOptions.anonrouter. This model selection requires a TEE route and allows one provider attempt.

import { anonrouter } from '@anonrouter/ai-sdk-provider';

const model = anonrouter('deepseek/deepseek-v3.2', {
  provider: { minimum_privacy: 'tee', allow_fallbacks: false },
});

Read result.providerMetadata?.anonrouter for the selected model, provider, privacy tier, and cost when the API returns them. See provider routing for policy fields and reasoning controls for model-specific effort settings.

For managed web search, choose an eligible model and set webSearch: { maxUses: 1 } when you select it. The adapter exposes the search summary in providerMetadata.anonrouter.webSearch. See web search for eligible routes, limits, and pricing.

Compatibility mode

For an existing bearer-key integration or cross-model fallback, create the provider with mode: 'compatibility' and a compatibility-enabled inference key.

import { createAnonRouter } from '@anonrouter/ai-sdk-provider';

const anonrouter = createAnonRouter({ mode: 'compatibility' });

In this mode the broker receives the API-key identity and request content together in memory. Pass models as an ordered list of up to four fallback model IDs if you need cross-model fallback. See compatibility mode for supported endpoints and key settings.

For independent attestation or end-to-end encryption, use the confidential SDK.

Troubleshooting

401 or 403. Confirm that the server has an inference key. If you chose compatibility mode, confirm that OpenAI-compatible access is enabled for it.
Unknown model. Copy a current model ID from the catalog.
Tool calls fail. Check the selected route's tool support and your tool schema.

See the npm package for factory settings and additional examples.

Vercel AI SDK