Integration Guides

Works with

Vercel AI SDK

Use ai@6 with @ai-sdk/openai-compatible@2 for Chat Completions model calls through Keel. Keep the Keel key on the server and configure the provider key in Keel. The fetch adapter adds the timestamp and nonce required for each HTTP attempt, then sends the SDK call to Keel's public proxy route. The invocation-level headers preserve one idempotency key across retries of that model request. Save its exact prompt and key for an unchanged retry; create a new key for each new invocation. Do not use one global key across prompts or tool-loop steps. Do not switch to a new key after an unknown outcome; reconcile the original request first.

npm install ai@6 @ai-sdk/openai-compatible@2

import { generateText } from "ai";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";

const keel = createOpenAICompatible({
  name: "keel",
  baseURL: "https://api.keelapi.com/v1/proxy/openai",
  apiKey: process.env.KEEL_API_KEY!,
  fetch: (input, init) => {
    const proxyEndpoint = "https://api.keelapi.com/v1/proxy/openai";
    const endpoint = proxyEndpoint + "/chat/completions";
    const target = input instanceof Request ? input.url : String(input);
    if (target !== endpoint) throw new TypeError("SDK request is outside the Keel chat endpoint");
    const headers = new Headers(input instanceof Request ? input.headers : undefined);
    new Headers(init?.headers).forEach((value, name) => headers.set(name, value));
    headers.set("X-Keel-Timestamp", String(Math.floor(Date.now() / 1000)));
    headers.set("X-Keel-Nonce", "nonce-" + crypto.randomUUID().replaceAll("-", ""));
    const outbound = input instanceof Request ? new Request(proxyEndpoint, input) : proxyEndpoint;
    return fetch(outbound, { ...init, headers });
  },
});

const idempotencyKey = crypto.randomUUID(); // Save with this invocation for an unchanged retry.
const { text } = await generateText({
  model: keel.chatModel("gpt-4o-mini"),
  prompt: "Hello",
  headers: { "Idempotency-Key": idempotencyKey },
});

Before the first call, enable the proxy route for your Keel project, allow the chosen model in policy, and test the server-side OpenAI provider key in Keel. If you see request_not_fresh, check the timestamp and nonce adapter; proxy_routes_disabled means the project has the compatibility route off; upstream_error requires checking the provider configuration or response.

This routes model calls through Keel. Requests that include tool definitions are refused before dispatch (platform_safety_floor.unknown_action); to govern tools your application executes, route those actions through Keel separately, such as its managed MCP call path. See theintegration coverage guidefor the exact boundary of each stack item.