neon-ai-gateway — скилл для Claude Code: один API ко всем LLM

---
name: neon-ai-gateway
description: >-
  One API and one credential for frontier and open-source LLMs and
  embedding models, built into your Neon branch and powered by Databricks.
  Use when a user wants to call an LLM, generate embeddings, add AI/chat/an
  agent to their app, route between model providers (OpenAI, Anthropic,
  Google/Gemini, Meta, Alibaba, and more), or avoid juggling separate
  provider API keys and accounts — especially when they already use Neon
  and want AI requests to branch with their project. Works with the OpenAI
  SDK, Anthropic SDK, google-genai, the Vercel AI SDK, and Mastra by
  changing only the base URL. Triggers include "call an LLM",
  "add AI to my app", "chat completion", "model routing", "LLM proxy/gateway",
  "one API for all
  models", "use Claude/GPT/Gemini", "AI SDK", "Mastra agent", "Neon AI
  Gateway", "generate embeddings", "generate images", and "log/rate-limit AI
  calls".
metadata:
  parent: neon
  source: https://github.com/neondatabase/agent-skills/tree/main/skills/neon-ai-gateway
---

**FIRST**: Use the parent `neon` skill for a Neon overview, getting started with Neon, Neon development best practices, and more.

If the `neon` skill is not installed, fetch it from https://neon.com/docs/ai/skills/neon/SKILL.md or install it with:

```bash
neon skills -s neon -y
```

# Neon AI Gateway

Currently available in `aws-us-east-2`, `aws-us-east-1`, `aws-eu-central-1`, and `aws-ap-southeast-1`.

The Neon AI Gateway brings LLM and embedding inference to your Neon branch. One API and one Neon credential give you access to frontier and open-source models from many providers (Anthropic, OpenAI, Google, Meta, and more), plus embedding models, all hosted and powered by Databricks. The catalog shifts over time, so treat `/v1/models` and the [models.dev Neon page](https://models.dev/providers/neon) as the source of truth rather than a fixed provider list. For embedding model IDs and dimensions, see the [embeddings guide](https://neon.com/docs/ai-gateway/embeddings). Your existing OpenAI/Anthropic/Gemini SDK works by changing only the base URL.

Use this skill to help the user send model calls through the gateway, wire it into the AI SDK or Mastra, and switch providers without rewiring code. Deliver a working inference request, a configured agent, or a precise answer from the official Neon docs.

## When to Use

Reach for the AI Gateway whenever an app or agent needs to call an LLM or generate text embeddings and the user would rather not manage model providers themselves:

- **One credential instead of many provider accounts.** A single Neon credential reaches the entire model catalog across every provider Databricks hosts. No separate OpenAI / Anthropic / Google billing, keys, or signups to provision and rotate.
- **Switch models without rewiring.** The unified endpoint is OpenAI-compatible and works with every model in the catalog — change one `model` field to move between Claude, GPT, and Gemini. Standard SDKs (OpenAI, Anthropic, google-genai) work with just a base-URL change.
- **AI follows your branches.** Each branch has its own gateway endpoint, scoped with the same lineage as your database. AI requests from a preview/feature branch are isolated to that branch — the same isolation your data already gets — which makes preview, CI, and agent environments self-contained.
- **No extra infrastructure, and it's already next to your data.** The gateway lives inside your Neon project (and is injected into Neon Functions automatically), runs on the same Databricks infrastructure that serves trillions of tokens a month, and supports streaming (SSE) out of the box.

If the user already has a deep, single-provider integration and no interest in Neon branching or multi-model routing, a direct provider SDK is fine — but the moment they want one credential, model portability, or branch-scoped AI, this is the reason to use it.

## What It Does

- **One API for all models** — Frontier and open-source models behind a single endpoint, addressed by their catalog ID (e.g. `claude-sonnet-4-6`, `gpt-5-mini`, `gemini-3-flash`).
- **Standard SDKs, one URL change** — OpenAI SDK and AI SDK (OpenAI-compatible MLflow/Responses routes), Anthropic SDK (native Messages), google-genai (native Gemini).
- **Branch-scoped** — Each branch gets its own gateway host; the Neon credential authorizes requests for that branch and its descendants.
- **Streaming** — Chat generation supports server-sent events with no extra configuration.
- **Embeddings** — Generate vectors for semantic search.

## Availability

Check these preconditions before setting anything up:

The AI Gateway is currently available in `aws-us-east-2`, `aws-us-east-1`, `aws-eu-central-1`, and `aws-ap-southeast-1`. Confirm the user's project is in one of these regions.

The AI Gateway is available on paid Neon plans only (Launch and Scale) and draws down prepaid AI Gateway credits. Any paid account with prepaid credits can call every model in the catalog. See https://neon.com/docs/ai-gateway/prepaid-credits.md.

### Enabling the gateway: plan and model access

The AI Gateway is credential-gated rather than a provisioning step. The CLI surfaces both plan and model-access limits:

- **Free plan → provisioning is blocked.** `neon config apply` / `deploy` and `neon checkout` **refuse** to enable the gateway on a Free plan (the gateway can't serve requests there), with a friendly "upgrade to a paid plan, or remove `aiGateway`" error. A dry-run `neon config plan` and `neon env pull` don't provision, so they only **warn**. So: to use the gateway the project's account must be on a paid Neon plan.
- **Paid plan with locked models.** `GET /v1/models` lists every catalog model and marks the ones the account can't call yet with `enabled: false`. `neon env pull` (and the env pull bundled into `apply` / `deploy` / `checkout`) warns when some models are enabled but no flagship model (e.g. Anthropic Opus, OpenAI Codex / `*-pro`) is among them, and links to the branch's AI Gateway page in the Neon Console (`https://console.neon.tech/app/projects/<project-id>/branches/<branch-id>/ai-gateway`). That page lists locked models under **Some models are restricted**, grouped by reason, each with its fix: **Upgrade plan** for a paid plan, **Add credits** for prepaid credits, or Support for an account restriction or a model not available to the account. After the user upgrades or buys credits, models unlock once Neon confirms access. No CLI warning does not mean every model is enabled; check the model's `enabled` value. See https://neon.com/docs/ai-gateway/troubleshooting.md.

## Setup

The gateway is part of `neon.ts` (see the `neon` skill for the branch-first workflow and `neon.ts` basics). Enable it with `aiGateway`:

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

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

```bash
neon deploy   # provisions the gateway on the linked branch
```

## Neon Infrastructure as Code (`neon.ts`)

The `aiGateway` toggle above is part of `neon.ts`, Neon's infrastructure-as-code file — one TypeScript file declares the gateway alongside every other branch service, in version control (see the `neon` skill for the full reference). Reconcile it against a branch the Terraform way:

```bash
neon config status   # print the branch's live config (is the gateway on?)
neon config plan     # dry-run diff of what apply would change
neon config apply    # enable the gateway on the branch  (neon deploy is an alias)
```

The gateway is **branch-scoped**: each branch gets its own gateway host. When a `neon.ts` is present, `neon checkout` applies the policy as it _creates_ a branch, so a fresh preview/CI branch comes up with the gateway already enabled. Checking out an _existing_ branch doesn't reconcile it — run `neon deploy` to apply changes. Provisioning (`config apply` / `deploy`), `link`, and `checkout` also pull the branch's gateway credentials into your local `.env.local`, so local runs hit the same branch gateway as the deployed function (no manual `env pull` needed).

## Environment Variables

When `aiGateway` is enabled, Neon injects the gateway credentials as **Neon-branded** env vars. Inside a deployed Neon Function these are injected automatically; locally, `neon env pull` writes them to `.env`/`.env.local` (or use `neon-env run -- <cmd>` to inject at runtime without a file):

| Variable                   | Meaning                                                                                                                             |
| -------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- |
| `NEON_AI_GATEWAY_TOKEN`    | Gateway bearer token (a Neon credential, `nt_live_...`)                                                                             |
| `NEON_AI_GATEWAY_BASE_URL` | **Bare branch gateway host** (`scheme://host`, **no path** — no `/ai-gateway`): `https://<branch-id>-api.ai.<region>.aws.neon.tech` |

> Neon injects **only** these two vars — it does **not** set `OPENAI_API_KEY` / `OPENAI_BASE_URL`. The `@neon/ai-sdk-provider` and Mastra's `neon/<model>` read `NEON_AI_GATEWAY_*` directly (zero config); for the plain OpenAI SDK / `@ai-sdk/openai`, build the client's `apiKey` + `baseURL` from them (shown below), or set your own `OPENAI_*` by hand (`env pull` leaves user-set vars untouched).

`NEON_AI_GATEWAY_BASE_URL` is the **bare host** — you append the dialect path yourself (which is exactly what the `@neon/ai-sdk-provider` does for you). The routes under the host are:

- `/v1` — unified, OpenAI **Chat Completions**-compatible; recommended default, works with every provider (`/v1/chat/completions`).
- `/openai/v1` — OpenAI **Responses** API (required for `gpt-5-…-codex` variants and `gpt-5-5-pro`); the `@ai-sdk/openai` provider uses the Responses API by default (`/openai/v1/responses`).
- `/anthropic` — native Anthropic Messages (extended thinking, prompt caching). Give the Anthropic SDK this as its base URL and it appends `/v1/messages` itself, so the full request path is `/anthropic/v1/messages`.
- `/gemini` — native Gemini `generateContent`. Give google-genai this as its base URL and it appends `/v1beta/models/<model>:generateContent` itself, so the full request path is `/gemini/v1beta/models/<model>:generateContent`.

So `${NEON_AI_GATEWAY_BASE_URL}/v1` is the chat-completions endpoint and `${NEON_AI_GATEWAY_BASE_URL}/openai/v1` the OpenAI Responses endpoint (both appended by you); for the native Anthropic and Gemini dialects you hand the SDK the shorter `/anthropic` or `/gemini` base and it appends the rest. See [Use with Plain SDKs](#use-with-plain-sdks-lower-level) below.

For typed, validated access to the injected credentials, pass the same `neon.ts` config object to `parseEnv` from `@neon/env` — it returns an `env.aiGateway` namespace (`apiKey`, `baseUrl`) derived from your config.

## Build Agents with the Vercel AI SDK (Recommended)

The [Vercel AI SDK](https://ai-sdk.dev) is the recommended way to call the gateway and build agents from TypeScript: one set of primitives (`generateText`, `streamText`, tool calling, structured output) over every catalog model, with first-class streaming for the long agent responses Neon Functions are built to host.

The dedicated `@neon/ai-sdk-provider` reads `NEON_AI_GATEWAY_BASE_URL` + `NEON_AI_GATEWAY_TOKEN` from the injected env with **zero config** and routes each model to the best endpoint (Anthropic → Messages, OpenAI/Codex → Responses, everything else → MLflow). On a Neon Function that streams text and generates images, pick an OpenAI catalog model, since the image tool runs on the Responses route:

```typescript
import { neon } from "@neon/ai-sdk-provider";
import { streamText } from "ai";

const result = streamText({
  model: neon("gpt-5-mini"),
  messages,
  tools: {
    image_generation: neon.tools.imageGeneration({
      model: "gpt-image-2",
      outputFormat: "jpeg",
      size: "1024x1024",
    }),
  },
});
return result.toUIMessageStreamResponse();
```

A single completion is the same provider with `generateText`:

```typescript
import { neon } from "@neon/ai-sdk-provider";
import { generateText } from "ai";

const { text } = await generateText({
  model: neon("claude-haiku-4-5"), // or gpt-5-3-codex, gemini-3-flash, ...
  prompt: "Summarize Postgres for me.",
});
```

> Prefer `@neon/ai-sdk-provider` over the bare `@ai-sdk/openai` `openai()`: Neon injects only `NEON_AI_GATEWAY_*`, not `OPENAI_*`, so `openai()` won't pick up the gateway from the env on its own. If you do use `@ai-sdk/openai`, configure it explicitly with `createOpenAI({ apiKey: process.env.NEON_AI_GATEWAY_TOKEN, baseURL: `${process.env.NEON_AI_GATEWAY_BASE_URL}/openai/v1` })`.

To build an **agent** — a model that calls tools in a loop and then answers — add `tools` and a `stopWhen` budget. The loop runs in-process, so on a Neon Function it isn't cut off by lambda-style timeouts:

```typescript
import { neon } from "@neon/ai-sdk-provider";
import { generateText, tool, stepCountIs } from "ai";
import { z } from "zod";

const { text } = await generateText({
  model: neon("claude-sonnet-4-6"),
  prompt: "How many open todos do I have, and what's the oldest one?",
  tools: {
    listTodos: tool({
      description: "List the user's open todos.",
      inputSchema: z.object({}), // AI SDK v5+: `inputSchema`, not `parameters`
      execute: async () => db.select().from(todos),
    }),
  },
  stopWhen: stepCountIs(5), // let the model call tools, then summarize
});
```

For a full AI SDK agent deployed as a Neon Function (streaming, tool calling, image generation, persistence), see the `neon-functions` skill's [references/ai-sdk.md](https://neon.com/docs/ai/skills/neon-functions/references/ai-sdk.md).

## Generate Images

Images come from the OpenAI Responses `image_generation` built-in tool on an OpenAI model, on the `/openai/v1/responses` route. The gateway has no image model endpoint, so the AI SDK's `generateImage()` throws `NoSuchModelError`. Every image request follows two rules:

- **Use a supported image model.** The gateway supports `gpt-image-2`, `gpt-image-1.5`, and `gpt-image-1-mini`. Set `model` explicitly on every tool definition, and use `gpt-image-2` unless the user asks for another.
- **Always stream.** A non-streaming response whose image is over 655,360 bytes fails with `400: The response content length cannot be greater than 655360 bytes long`, and most images are larger than that. Use `streamText`, never `generateText`, when the image tool is attached. With the plain OpenAI SDK, pass `stream: true` to `client.responses.create()`.

```typescript
import { neon } from "@neon/ai-sdk-provider";
import { streamText } from "ai";

const result = streamText({
  model: neon("gpt-5-mini"),
  prompt: "Generate an image of a red apple on a wooden table",
  tools: {
    image_generation: neon.tools.imageGeneration({ model: "gpt-image-2" }),
  },
});

for await (const part of result.fullStream) {
  if (
    part.type === "tool-result" &&
    !part.dynamic &&
    "result" in part.output &&
    typeof part.output.result === "string"
  ) {
    const png = Buffer.from(part.output.result, "base64"); // save or upload it
  }
}
```

## Generate Embeddings with the AI SDK

Use `@neon/ai-sdk-provider` 1.0.0+ with AI SDK 6+ for embedding calls. `neon.embeddingModel()` reads the same injected gateway variables as the language models:

```typescript
import { neon } from "@neon/ai-sdk-provider";
import { embed, embedMany } from "ai";

const model = neon.embeddingModel("qwen3-embedding-0-6b");
const providerOptions = { neon: { dimensions: 1024 } };
const { embeddings } = await embedMany({
  model,
  values: documentBodies, // unprefixed document text
  providerOptions,
  maxParallelCalls: 1,
});
const { embedding } = await embed({
  model,
  value: `Instruct: Given a question about Neon, retrieve documentation passages that answer the question\nQuery: ${queryText}`,
  providerOptions,
});
```

The query instruction is specific to Qwen3 and Neon documentation search; adapt it to the corpus. Documents stay unprefixed. Keep the same model and dimensions for documents and queries, and keep the original query text for BM25. For schema, storage, batch limits, and an OpenAI SDK alternative, follow the `neon-postgres` skill's [vector search guide](https://neon.com/docs/ai/skills/neon-postgres/references/vector-search.md).

## Build Agents with Mastra (Recommended)

[Mastra](https://mastra.ai) is the recommended framework when you want batteries-included agents — built-in memory, tools, workflows, and tracing — with the model still pointed at the gateway. With `@mastra/core` 1.47+, use a `neon/<model>` magic string; Mastra reads `NEON_AI_GATEWAY_BASE_URL` and `NEON_AI_GATEWAY_TOKEN` from the environment (injected by `neon deploy` when `aiGateway` is enabled). Use `parseEnv` only for other declared services (e.g. `env.postgres.databaseUrl` for `@mastra/pg` memory):

```typescript
import { Agent } from "@mastra/core/agent";
import { parseEnv } from "@neon/env";
import config from "../neon";

const env = parseEnv(config);

export const personalAssistant = new Agent({
  id: "personal-assistant",
  name: "personal-assistant",
  instructions:
    "You are a warm, concise personal assistant with long-term memory.",
  model: "neon/claude-haiku-4-5",
  memory, // your Mastra memory store, e.g. @mastra/pg on env.postgres.databaseUrl
});
```

## Use with Plain SDKs (Lower-Level)

When you don't need an agent framework — a single completion, an existing provider-SDK integration, or native provider features — call the gateway with the plain SDKs. Neon injects the `NEON_AI_GATEWAY_*` vars (not `OPENAI_*`), so set the client's `apiKey` + `baseURL` from them. For the OpenAI **Responses** dialect (`/openai/v1`):

```typescript
import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.NEON_AI_GATEWAY_TOKEN,
  baseURL: `${process.env.NEON_AI_GATEWAY_BASE_URL}/openai/v1`,
});

const res = await client.responses.create({
  model: "gpt-5-mini", // swap to claude-sonnet-4-6, gemini-3-flash, ...
  input: "What is Neon?",
});
```

For the unified **chat-completions** dialect, point `baseURL` at `/v1` instead:

```typescript
const client = new OpenAI({
  apiKey: process.env.NEON_AI_GATEWAY_TOKEN,
  baseURL: `${process.env.NEON_AI_GATEWAY_BASE_URL}/v1`,
});

const res = await client.chat.completions.create({
  model: "claude-sonnet-4-6",
  messages: [{ role: "user", content: "What is Neon?" }],
});
```

The Anthropic SDK and google-genai work the same way for native provider features — point the Anthropic SDK at `${NEON_AI_GATEWAY_BASE_URL}/anthropic` (it appends `/v1/messages` itself) and google-genai at `${NEON_AI_GATEWAY_BASE_URL}/gemini` (it appends `/v1beta/models/...`).

## Model Identifiers

Use a model's catalog ID directly in the `model` field — e.g. `claude-sonnet-4-6`, `gpt-5-mini`, `gemini-3-flash`. No provider prefix is needed. To look up the exact identifiers the gateway serves, which underlying model each maps to, and their context windows, pricing, and capabilities, use any of:

- **models.dev Neon provider page: https://models.dev/providers/neon** — the canonical, always-current list of the Neon provider's model IDs and their underlying models. The machine-readable catalog is at https://models.dev/api.json (the `neon` key).
- **Models doc:** see Further Reading.

## List Available Models at Runtime (`/v1/models`)

The gateway also exposes the model catalog **live from your own branch endpoint**, so an app or agent can discover exactly which models this branch serves without hard-coding the list. It is an OpenAI-compatible list endpoint, served **only on the unified dialect** (`/v1`):

```bash
curl "$NEON_AI_GATEWAY_BASE_URL/v1/models" \
  -H "Authorization: Bearer $NEON_AI_GATEWAY_TOKEN"
```

- `GET ${NEON_AI_GATEWAY_BASE_URL}/v1/models` → **200**
- `GET ${NEON_AI_GATEWAY_BASE_URL}/openai/v1/models` → **404** (not served on the Responses dialect — use `/v1`)

**Getting the credentials for the request.** Both values come from the same branch-scoped Neon credential the gateway uses everywhere else — you never manage a provider key:

- **Provision via `neon.ts` (recommended).** Enable `aiGateway` in `neon.ts` and run `neon deploy` (or `neon config apply`). Provisioning, `neon link`, and `neon checkout` pull `NEON_AI_GATEWAY_TOKEN` + `NEON_AI_GATEWAY_BASE_URL` into your local `.env.local`; inside a deployed Neon Function they're injected automatically. See **Setup** and **Environment Variables** above.
- **Pull into the environment via CLI.** `neon env pull` writes the two vars to `.env`/`.env.local`, or `neon-env run -- <cmd>` injects them at runtime without a file — but only when `neon.ts` declares `aiGateway`; the vars are never pulled off branch state alone.
- **Provision via the Console UI.** Enable the AI Gateway on the branch in the Neon Console and copy the branch's gateway base URL and a Neon credential (token) from the project's connection/credentials view.

Any Neon credential (`nt_live_...`) valid for the branch works as the bearer token; `NEON_AI_GATEWAY_BASE_URL` is the bare branch host (no path).

**Response shape** — OpenAI/OpenRouter-compatible list:

<!-- prettier-ignore -->
```jsonc
{
  "object": "list",
  "data": [
    {
      "id": "claude-sonnet-4-6",              // catalog model ID — use directly in the `model` field
      "canonical_slug": "claude-sonnet-4-6",
      "name": "Claude Sonnet 4.6",            // human-readable display name
      "object": "model",
      "owned_by": "anthropic",                // provider slug, e.g. anthropic | openai | google | meta | alibaba | databricks | ... (non-exhaustive; read live)
      "created": 0,
      "enabled": true,
      "context_length": null,
      "architecture": {
        "modality": "text->text",
        "input_modalities": ["text"],
        "output_modalities": ["text"],
        "tokenizer": "Claude",                // Claude | Gemini | GPT | "" (empty for open-source)
        "instruct_type": null
      },
      "top_provider": {
        "is_moderated": false,
        "context_length": null,
        "max_completion_tokens": null
      },
      "pricing": null,
      "per_request_limits": null
    }
    // ... one entry per model in the branch's catalog
  ]
}
```

> Note: `context_length`, `pricing`, and `per_request_limits` are currently `null` and `created` is `0` for every entry — for context windows, pricing, and capabilities use the models.dev catalog above. Use `/v1/models` when you need the live, branch-scoped list of servable model IDs (e.g. to populate a model picker or validate a `model` before a request), and keep only entries with `enabled: true`.

## Neon Documentation

The Neon documentation is the source of truth and the AI Gateway is evolving rapidly, so always verify against the official docs. Any doc page can be fetched as markdown by appending `.md` to the URL or by requesting `Accept: text/markdown`. Find the right page from the docs index (https://neon.com/docs/llms.txt) and the changelog announcements.

## Further Reading

- https://neon.com/docs/ai-gateway/overview.md
- https://neon.com/docs/ai-gateway/get-started.md
- https://neon.com/docs/ai-gateway/models.md
- https://neon.com/docs/ai-gateway/chat-completions.md
- https://neon.com/docs/ai-gateway/anthropic-messages.md
- https://neon.com/docs/ai-gateway/openai-responses.md
- https://neon.com/docs/ai-gateway/gemini.md
- https://neon.com/docs/ai-gateway/authentication.md
- https://neon.com/docs/ai-gateway/troubleshooting.md