Route LLM calls through Neon's branch-scoped AI Gateway

Helps a user call LLMs through the Neon AI Gateway, a preview LLM inference layer built into a Neon branch that fronts frontier and open-source models from seven providers - Anthropic, OpenAI, Google, Meta, Alibaba, DeepSeek, and Databricks (the gateway itself runs on Databricks infrastructure) - behind one API and one Neon credential, with existing OpenAI/Anthropic/Gemini SDKs working after only a base-URL change. The gateway is enabled per-branch via neon.ts's preview.aiGateway toggle, reconciled with neon config status/plan/apply (or the neon deploy alias) the same infrastructure-as-code way as other branch services; a fresh branch created via neon checkout with neon.ts present comes up with the gateway already on, while checking out an existing branch does not reconcile it. Provisioning injects OpenAI-standard env vars (OPENAI_API_KEY, OPENAI_BASE_URL - which already includes the /ai-gateway/openai/v1 Responses-API path) plus Neon-branded aliases (NEON_AI_GATEWAY_TOKEN, NEON_AI_GATEWAY_BASE_URL - the bare host with no path), or parseEnv from @neon/env for typed access via env.aiGateway. Four dialect routes sit under the gateway host: /ai-gateway/mlflow/v1 (unified OpenAI Chat Completions-compatible, the recommended default across all providers), /ai-gateway/openai/v1 (OpenAI Responses API, required for the gpt-5-...-codex variants and gpt-5-5-pro), /ai-gateway/anthropic/v1 (native Anthropic Messages), and /ai-gateway/gemini/v1beta (native Gemini generateContent). The skill helps build agents with the recommended frameworks: the Vercel AI SDK's generateText/streamText plus tool calling and a stopWhen step budget (running in-process so it isn't cut off by lambda-style timeouts), either via @ai-sdk/openai reading the injected OpenAI env vars automatically or @neon/ai-sdk-provider for multi-provider routing that sends Anthropic models to Messages, OpenAI/Codex to Responses, and everything else to MLflow; or Mastra for memory-backed agents (threads/messages in Postgres via @mastra/pg) pointed at the MLflow dialect through parseEnv. Plain SDK usage is also covered - new OpenAI() picks up the injected env with zero config for the Responses dialect, and swapping /openai/v1 for /mlflow/v1 on the base URL switches to chat completions. Model IDs are used directly with no provider prefix (e.g. claude-sonnet-4-6, gpt-5-mini, gemini-2-5-flash), looked up via the models.dev Neon provider page or its machine-readable api.json catalog.

// neon.ts
import { defineConfig } from "@neon/config/v1";

export default defineConfig({
  preview: {
    aiGateway: true,
  },
});

Use it whenever an app or agent needs to call an LLM and the user wants one credential instead of managing separate provider accounts, model portability without rewiring code, or branch-scoped AI requests that follow the same isolation as their Neon data. If the user already has a deep single-provider integration with no interest in branching or multi-model routing, a direct provider SDK is fine instead. The gateway is a preview feature available only on new projects in the us-east-2 region - it cannot be enabled on existing projects - and foundation model access requires a paid Neon plan; users without access should be pointed to the private beta sign-up.

Input is a chat/completion request (prompt, messages, or an agent's tool-calling loop) addressed to a catalog model ID. Output is a standard OpenAI-, Anthropic-, or Gemini-shaped response, with streaming (server-sent events) supported on every endpoint with no extra configuration.

Injects OpenAI-standard and Neon-branded environment variables automatically inside a deployed Neon Function (or via neon env pull / neon-env run locally), and integrates with the OpenAI SDK, Anthropic SDK, google-genai, the Vercel AI SDK (directly or via @neon/ai-sdk-provider), and Mastra. The Neon documentation is called out as the source of truth given the feature's rapid evolution, with any doc page fetchable as markdown by appending .md to its URL.

Developers building AI features or agents on Neon who want a single credential and endpoint across multiple model providers, branch-scoped AI environments for preview/CI/agent workflows, and a recommended path through either the Vercel AI SDK or Mastra rather than wiring up each provider's SDK separately.