@eidentic/nextjs

Next.js App Router integration for Eidentic — withEidentic wraps an agent as a POST route handler that emits an AI SDK UI-compatible stream, and eidenticNextConfig patches your next.config.ts to prevent native-addon bundling errors (better-sqlite3, etc.). Works with useChat from the Vercel AI SDK out of the box.

Install

pnpm add @eidentic/nextjs @eidentic/libsql

Use @eidentic/libsql (pure-JS) instead of @eidentic/sqlite (native addon) in Next.js and other bundler environments.

Usage

// next.config.ts
import { eidenticNextConfig } from "@eidentic/nextjs";
import type { NextConfig } from "next";

export default eidenticNextConfig({} satisfies NextConfig);

// app/api/chat/route.ts
import { withEidentic } from "@eidentic/nextjs";
import { Agent, AIModel } from "eidentic";
import { LibsqlStore } from "@eidentic/libsql";
import { anthropic } from "@ai-sdk/anthropic";

export const runtime = "nodejs";

const agent = new Agent({
  id: "support",
  instructions: "You are a helpful support assistant.",
  model: new AIModel(anthropic("claude-sonnet-4-5")),
  store: new LibsqlStore("file:eidentic.db"),
});

export const POST = withEidentic(agent, {
  async identify(req) {
    const session = await requireSession(req);
    return { userId: session.user.id, orgId: session.org.id };
  },
});

// Client component — Vercel AI SDK v7+
import { useChat } from "@ai-sdk/react";

export function Chat() {
  // `useChat` POSTs `{ messages: [...] }`; `withEidentic` reads the newest user message out of the
  // box — no `prepareSendMessagesRequest` / request transform needed. The AI SDK's top-level chat
  // `id` becomes the stable Eidentic session id unless an explicit `sessionId` is supplied.
  const { messages, sendMessage } = useChat();
  return null; // render `messages`; send with `sendMessage({ text })`
}

Request-body userId, orgId, and apiKey values are ignored by default. Identity must come from the trusted identify(req) hook; omitting it returns HTTP 401. allowUntrustedIdentityBody: true restores spoofable body identity, while unsafeAllowAnonymous: true restores ownerless sessions. Both options exist only for a controlled single-tenant migration.

AI SDK regenerate-message requests return HTTP 409. Eidentic sessions are append-only; replaying an old user message into the same session would retain the stale assistant answer and duplicate history. Start a new/forked chat id when implementing regeneration.

Links

• GitHub
• Issue tracker
• Root README

Apache-2.0