@braingrid/ai-sdk-provider

Vercel AI SDK provider for the BrainGrid AI Gateway. Apps built on BrainGrid call Anthropic & Gemini through the gateway with no LLM account, no API keys, and no auth code — the project's signing key is provisioned automatically, and the provider mints short-lived, per-session tokens on the server.

It's a thin wrapper over @ai-sdk/anthropic: the gateway speaks the Anthropic Messages shape and translates other providers (Gemini, etc.) server-side, so a single provider serves every model your plan allows.

Install

npm i @braingrid/ai-sdk-provider ai

Usage

import { generateText, streamText } from 'ai';
import { braingrid } from '@braingrid/ai-sdk-provider';

const ai = braingrid(); // reads BRAINGRID_PROJECT_SECRET_KEY from env — no setup

// Claude
const { text } = await generateText({
  model: ai('claude-sonnet-4-6'),
  prompt: 'Summarize this support ticket: …',
});

// Gemini — same call, different model string
const { textStream } = await streamText({
  model: ai('gemini-2.5-pro'),
  prompt: 'Write a release note for v2.3 …',
});

Everything else — tool calls, structured outputs, multi-modal, streaming — comes from the AI SDK itself. There is no BrainGrid-specific API to learn.

Attribution

Carry per-end-user attribution into the gateway's usage events (for reporting, never for billing identity):

const ai = braingrid({
  attribution: { end_user_id: 'cust_8842', metadata: { plan: 'pro' } },
});

How it works

BRAINGRID_PROJECT_SECRET_KEY is a single bgpk_… credential bundling the project's private signing key, its key id, and the gateway URL. braingrid() decodes it and, for each request, mints a short-lived JWT (30-minute TTL, cached and re-minted on expiry) signed with that key. The gateway resolves the billing org/project from the key id, checks the model against the project's plan allowlist, proxies to the provider, and bills BrainGrid credits.

Server-only. The secret key must never reach the browser — call braingrid() from server code (route handlers, server actions), exactly as you would a provider holding an API key.

Direct (non-AI-SDK) calls

If you call the gateway HTTP API yourself, mint a session token by hand:

import { mintSession } from '@braingrid/ai-sdk-provider';

const token = await mintSession({ attribution: { end_user_id: 'cust_8842' } });
// fetch(`${gatewayUrl}/v1/messages`, { headers: { Authorization: `Bearer ${token}` }, … })

License

Apache-2.0