@samirdamle/dci-server

Helpers for DCI endpoints: parse and validate the request, stream the answer as Server-Sent Events, and always finish the response. Web-standard, so it runs on Node, Bun, Deno, Next.js route handlers and edge runtimes. No LLM SDK is bundled: calling a model is your code.

npm install @samirdamle/dci-server

import { dciHandler, formatContextForPrompt } from '@samirdamle/dci-server';

export const POST = dciHandler(async (req, stream, { signal }) => {
  const context = formatContextForPrompt(req.context); // the selected items, for your prompt
  stream.toolStart('t1', 'lookup', 'Looking it up…');
  stream.toolEnd('t1', true);
  stream.text(`You asked "${req.prompt}" about:\n${context}`); // or stream your model's tokens
  if (signal.aborted) return; // the user pressed Stop
  stream.clientAction('highlight', { ids: req.context.flatMap((n) => (n.id ? [n.id] : [])) });
});

For Node's http module or Express, wrap it with toNodeHandler from @samirdamle/dci-server/node.

• A backend with Claude in ~30 lines
• The protocol
• Examples

MIT licensed.