Turn on Jev

The Mock Provider only understands exact names. Jev, TypeSafe's typed-decision model, understands the wording people actually use: "who manages the account", "biggest first", "also group by advisor". GridCue calls Jev from your server, never from the browser, so your key stays private.

Two changes: the browser talks to your endpoint instead of the Mock, and your server answers that endpoint with Jev.

1. Install the TypeSafe SDK

npm i gridcue @typesafe-ai/sdk

@typesafe-ai/sdk is an optional peer dependency of gridcue. Only gridcue/server uses it.

2. Point the browser at your endpoint

Replace createMockProvider() with createRemoteProvider, from the root gridcue entry. It posts each request to a URL you choose:

import { createGridCue, createRemoteProvider } from "gridcue";

const cue = createGridCue({
  schema,
  adapter: createTanStackAdapter({ schema, table }),
  provider: createRemoteProvider({ endpoint: "/api/gridcue" }),
});

createRemoteProvider also takes headers (for example, a CSRF token) and a custom fetch.

3. Mount the Server Handler

The Server Handler is a small endpoint that holds the provider's key and answers resolution requests. createGridCueHandler returns a Fetch-standard (request: Request) => Promise<Response> function. Give it a provider made by createJevProvider.

Next.js App Router:

// app/api/gridcue/route.ts
import { createGridCueHandler, createJevProvider } from "gridcue/server";

export const POST = createGridCueHandler({
  provider: createJevProvider({ apiKey: process.env.JEV_API_KEY }),
});

Express, or any Node (req, res) server, through toNodeHandler:

import express from "express";
import { createGridCueHandler, createJevProvider, toNodeHandler } from "gridcue/server";

const handler = createGridCueHandler({ provider: createJevProvider({ apiKey: process.env.JEV_API_KEY }) });

const app = express();
app.post("/api/gridcue", toNodeHandler(handler));

The handler accepts only POST with a JSON body up to 32 KB (maxBodyBytes), validates the request against the protocol, and never logs requests, payloads or keys. Other runtimes (Hono, Cloudflare Workers, Vite's dev server) are in Server Handler recipes.

Where the key goes

Put the key in a server-side environment variable named JEV_API_KEY. Locally, that is a git-ignored .env.local:

# .env.local
JEV_API_KEY=your-key

Only import gridcue/server from server code. In a browser build it throws on import, so a mistake fails loudly instead of shipping the key.

Choose a strategy

createJevProvider takes a strategy: "fan-out" (the default) asks more questions per part and handles compound, "also" and nested requests, while "focused" asks the fewest questions for simple, single-change requests. See Choosing a strategy.

createJevProvider({ apiKey: process.env.JEV_API_KEY, strategy: "focused" });

The provider pins the model to jev-1.13.0, the model GridCue's confidence bands are tuned against. Pass model to change it, and re-run your evals when you do.

Protect the endpoint

Every call to /api/gridcue spends your Jev credits, and GridCue adds no auth of its own. Put the endpoint behind the same session check and rate limits as the rest of your app before you ship. See Protect the endpoint.

Next

• Describe your domain: aliases, approved values, and domain nouns.
• Going to production: auth, rate limits, timeouts and audit.