@llm-ports/adapter-vercel

Vercel AI SDK adapter for llm-ports. Migration helper for users already using @ai-sdk/*.

Why this adapter exists

If you've already wired your project around @ai-sdk/anthropic, @ai-sdk/openai, etc., you can adopt llm-ports without rewriting that integration. This adapter takes your pre-configured Vercel LanguageModel instances and routes LLMPort calls through Vercel's generateText, streamText, embed, and embedMany helpers. You get cost gating, fallback chains, and capability factories on top of the stack you already have.

For new projects, prefer the direct adapters (@llm-ports/adapter-anthropic, @llm-ports/adapter-openai, @llm-ports/adapter-google) — fewer layers, more control.

Install

pnpm add @llm-ports/core @llm-ports/adapter-vercel ai @ai-sdk/anthropic

Configure

import { anthropic } from "@ai-sdk/anthropic";
import { openai } from "@ai-sdk/openai";
import { createRegistryFromEnv } from "@llm-ports/core";
import { createVercelAdapter } from "@llm-ports/adapter-vercel";

const registry = createRegistryFromEnv({
  adapters: {
    vercel: createVercelAdapter({
      models: {
        "claude-sonnet-4-6": anthropic("claude-sonnet-4-6"),
        "gpt-5": openai("gpt-5"),
      },
      embeddingModels: {
        "text-embedding-3-small": openai.textEmbeddingModel("text-embedding-3-small"),
      },
      // pricing is optional in alpha.8+ — bundled VERCEL_PRICING covers common
      // @ai-sdk/* models. Supply pricingOverrides only for missing entries.
    }),
  },
});

const llm = registry.getPort();

.env:

LLM_PROVIDER_FAST=vercel|claude-sonnet-4-6|cost:50/day
LLM_PROVIDER_GPT=vercel|gpt-5|cost:100/day
LLM_TASK_ROUTE_TRIAGE=fast,gpt

Adapter options

interface VercelAdapterOptions {
  models?: Record<string, LanguageModel>;             // pre-configured @ai-sdk/* instances
  embeddingModels?: Record<string, EmbeddingModel<string>>;
  pricing?: Record<string, ModelPricing>;             // optional; merges over VERCEL_PRICING
  validationStrategy?: ValidationStrategy;
  imageSizeLimitBytes?: number;                       // default 20 MB
  onRetry?: OnRetry;                                  // observability hook
}

Browser execution

This adapter does NOT expose its own dangerouslyAllowBrowser option, because the underlying provider client is constructed by you when you build the LanguageModel instance. For browser execution, pass the SDK's flag at LanguageModel construction time:

import { createOpenAI } from "@ai-sdk/openai";

const openai = createOpenAI({
  apiKey: ephemeralUserKey,
  dangerouslyAllowBrowser: true,        // ← set this on the @ai-sdk/* client
});

const adapter = createVercelAdapter({
  models: { default: openai("gpt-5") },
});

The same pattern applies to @ai-sdk/anthropic (and any other browser-restricted SDK). The adapter-openai and adapter-anthropic packages expose the flag directly because they construct the SDK client internally.

Bundled pricing

VERCEL_PRICING (alpha.8+) covers the common OpenAI / Anthropic / Google models used via @ai-sdk/*. Values mirror the direct adapters' bundled tables. For uncommon @ai-sdk/* providers (LMStudio, OpenRouter, perplexity-ai, custom routes), supply pricing entries yourself.

Supported features

| Feature | Status | |---------|--------| | generateText | ✓ | | generateStructured (Zod schemas) | ✓ (prompted JSON + retry-with-feedback + alpha.5 repair pass) | | streamText | ✓ | | streamStructured (partial JSON) | ✓ (best-effort partial parse) | | runAgent (multi-turn) | ✓ (alpha.8+; via Vercel's native tools + maxSteps loop) | | generateEmbedding / generateEmbeddings | ✓ | | Multimodal content blocks | ✓ (alpha.8+; via Vercel MessagePart[] shape) | | AbortSignal cancellation | ✓ entry + in-flight (alpha.6) |

Content blocks supported

text, image (base64 → data URI; URL → passthrough), audio (base64 → Vercel file part), tool_use, tool_result. Throws for URL-form audio (Vercel routes audio via file-data; pass base64 + mediaType instead).

Cancellation

Full AbortSignal support shipped in 0.1.0-alpha.6. The signal is passed through to Vercel's abortSignal field on generateText / streamText, so controller.abort() cancels the in-flight provider HTTP request (the cancellation propagates from Vercel to the underlying provider SDK). See the Cancellation guide.

Known limitations

• No reasoning-model headroom-multiplier. Unlike adapter-openai, the Vercel adapter doesn't apply a 10× headroom for OpenAI o-series / gpt-5-nano / Cerebras gpt-oss when called with small maxOutputTokens. The Vercel adapter does retry on finishReason: "length" with empty text + reasoning signal (alpha.1 fix #4), but it doesn't pre-seed the multiplier. Workaround: set maxOutputTokens 5-10× higher than your visible-output budget, or use @llm-ports/adapter-openai directly for reasoning models in v0.1.

When to use this vs the direct adapters

| Scenario | Choose | |---|---| | Already on @ai-sdk/* and want to add llm-ports | This adapter | | New project | adapter-anthropic / adapter-openai / adapter-google directly | | Need full multimodal richness (Anthropic prompt-caching, Gemini context-caching, etc.) | Direct adapters expose those; Vercel abstracts them away |

Reading next

• Vercel adapter docs — full feature deep-dive
• Migration from Vercel AI SDK guide — two migration paths (wrap vs. replace)
• Adapter feature matrix — direct adapter comparison