Middleware

Wrap every tool call: log it, block it, or change what the AI gets back. This needs an AI that can use tools.

import { defineTool } from "@leuria/client"

const findCustomer = defineTool({
  name: "find_customer",
  description: "Find a customer by name. Returns their name, city and email.",
  inputSchema: { type: "object", properties: { name: { type: "string" } }, required: ["name"] },
  execute: ({ name }) => ({ name, city: "Lyon", email: "ada@example.com" }),
})

// Logs each call, and hides email addresses from the AI.
const redact = async (call, next) => {
  print(`→ ${call.name} ${JSON.stringify(call.args)}`)
  const outcome = await next()
  if (!outcome.ok) return outcome
  const clean = JSON.parse(JSON.stringify(outcome.result).replace(/[\w.+-]+@[\w-]+\.[\w.]+/g, "[hidden]"))
  return { ok: true, result: clean }
}

const answer = await ai.chat({
  prompt: "Where does Ada live, and what is her email?",
  tools: [findCustomer],
  middleware: [redact],
  signal,
}).text()
print(answer)

How it works

• A middleware gets the call (name, args, callId, the turn's context) and next(), which runs the tool. It returns { ok: true, result } or { ok: false, error }.
• Put it on one request or conversation (middleware), or on every call of the page (createLeuria({ middleware })); the page-wide layers run first.
• Use it for audit logs, to block a call in some state, or to mark text you don't control as data, not instructions. See Page tools.