Serving over HTTP

@intellibrix/express wraps an Express app in a Brick and ships two ready-made routes.

import ExpressBrick from '@intellibrix/express'
import { Intelligence } from 'intellibrix'

const api = new ExpressBrick({
  name: 'api',
  port: 10100,
  host: '0.0.0.0', // the default is localhost; use 0.0.0.0 inside a container
  intelligence: new Intelligence({ provider, model, memory }),
})

api.askEndpoint() // POST /ask
api.askStreamEndpoint() // GET /ask/stream

POST /ask

curl -s localhost:10100/ask -H 'content-type: application/json' \
  -d '{"prompt":"What is 2+2?","threadId":"demo"}'
# {"reply":"4"}

A missing prompt returns 400; a failure returns 500 with { error }. threadId selects the conversation when the Intelligence has memory. If the client disconnects, the run is aborted so you do not pay for an answer nobody is waiting for.

GET /ask/stream

curl -N 'localhost:10100/ask/stream?prompt=Tell+me+a+story'

The response is Server-Sent Events, one event: <type> and data: <json> frame per stream event (text deltas, tool calls, the finish), followed by event: done. In the browser use EventSource. Disconnecting aborts the run here too.

Using the routes yourself

askRoute(target) and askStreamRoute(target) are plain Express handlers, and target can be a Brick or an Intelligence, so you can add them to an app you already have:

import { askRoute } from '@intellibrix/express'

app.post('/chat', express.json(), authenticate, askRoute(brick))

Before you expose it

• Authenticate. An open /ask lets anyone spend your tokens and use your tools. Add middleware (middleware: [authenticate] in the options runs before every route).
• Gate tools. There is no human in an HTTP request, so a tool that needsApproval is denied unless you provide an onToolApproval policy. Do not give a public endpoint tools that write.
• Limit cost. Put rate limits in front and keep maxSteps modest.
• Mind the body parser. ExpressBrick parses JSON by default; for signed webhooks see Working with GitHub.