Vercel AI Gateway

Install mayura and @mayurajs/gateway-vercel. The package uses the shared compatible transport from @mayurajs/provider-openai; its dependencies stay outside the core package.

import { createModels } from 'mayura';
import { vercelGateway } from '@mayurajs/gateway-vercel';

const gateway = vercelGateway({ apiKey: 'your-gateway-key' });
const models = createModels({ providers: [gateway], maxCallCostMicros: 50_000 });
// Set these example rates to conservative rates for your gateway model and routing configuration.
const model = models.model('vercel/anthropic/your-model', { pricing: {
  inputMicrosPerMillionTokens: 3_000_000,
  outputMicrosPerMillionTokens: 15_000_000,
} });
void model;

Credentials: Bearer gateway API key. No environment variables are read by the package. baseURL is an explicit HTTPS egress choice; credentials, query strings and fragments in URLs are refused. Never send gateway credentials to a browser. headers cannot replace the package's credential headers or transport headers.

Model calls need the runtime permission model:vercel/anthropic/your-model and a reserved call budget. Native Mayura tool permissions, authorizers and approvals still apply to any returned tool calls. A gateway cannot grant a tool. History and native tools use Chat Completions; strict structured output and streaming require a model and route that forward those features. output: 'json_object' explicitly selects JSON mode if needed. Set tokenLimitField: 'max_completion_tokens' when your model requires it. Media is disabled until the host explicitly sets media; support depends on the selected model and route.

Discovery: GET /models. gateway.discovery indicates whether this profile supports it. gateway.listModels({ permissions: { allow: ['gateway:vercel:models'] }, signal }) makes one explicit bounded request, with no pagination or retry, and returns names only. The response is limited to 1,000 models and 256 KiB, within the configured timeout. Listing does not install prices, populate the registry or grant model execution. Do not offer unfiltered discovery results to tenants as an authorization decision.

Billing uses the maximum of configured token pricing and usage.cost when the response reports it, rounded up to whole USD micros. LiteLLM also uses x-litellm-response-cost. A present malformed cost refuses the response. Known usage remains charged if output parsing fails or a stream ends after reporting usage. Token usage is required even when cost is reported. Optional host catalog prices are used only with prices: 'catalog'; this package ships no static gateway price list. Long-context host prices are supported.

Gateway routing, server-side retries, paid plugins and provider surcharges must be bounded by your gateway configuration and conservative host prices. Mayura adds no transport retry. A reported charge can exceed the pre-admitted bound; the runtime accounts it and fails the call, but cannot undo upstream spending. A dashboard estimate or a streaming header unavailable at completion is not a provider billing guarantee.

The common gateway guide describes scope, cost accounting and the owner-run qualification harness. Offline regression tests, builds and package checks passed locally. The paid harness has not been run; live gateway behavior remains unqualified. See release validation status.