One platform for voice, chat, and developers

Muwakil AI is not a standalone tool. It is an agent engine, a build studio, and an API, designed for the Arabic customers actually speak.

Customer Hi, did order 7952 arrive yet? What is the total?
Omar (preset voice) Welcome! Your order seven-nine-five-two went out for delivery just now, and the total is fifteen dinars on delivery.
JO

Jordanian number reading · No Gulf vocabulary bleed · Muwakil Arabic speech recognition

Voice agents

The agent hears the customer and replies with speech in the same call. The path is designed for barge-in, for hold phrases only when a tool is actually running, and for reading numbers and dates in the selected dialect.

  • Live call preview
  • ~0.37 s measured median reply latency and 20/20 consecutive-call stability
  • Voice cloning with the speaker’s consent

Chat agents

The same agent can reply in text, from your knowledge base, with tools, and without inventing facts you did not provide.

  • Per-agent knowledge base
  • Handoff to a human when needed
  • In-conversation memory, and longer memory if you enable it

Agent Studio

Muwakil Studio is Arabic-first, with English. Start from a template, set dialect, voice, and knowledge, then test before publish. A visual workflow canvas is part of the platform design.

  • Templates for stores, clinics, restaurants, real estate, and support
  • Draft, publish, and versions
  • Live test from the same board
1. Call Start Instant JO greeting Start · JO
2. Conversation Order ID & KB check Muwakil AI · KB
3. check_order Hold phrase on tool Webhook · #7952
4. Reply / Transfer Read total or handoff 15 JOD · End

API

For teams that already have a stack: a standard /v1 REST path, with isolated keys per product or customer.

  • /v1/chat/completions
  • /v1/audio/transcriptions and /v1/audio/speech
  • /v1/embeddings

Dialects: Jordanian, Saudi, Lebanese

You choose the dialect for the agent or for a conversation step. Vocabulary rules do not mix. Speech recognition is currently stronger on Levantine, so Jordanian and Lebanese are stronger today; Saudi understanding will improve when Gulf speech recognition support is added.

  • Jordanian (default and core product)
  • Saudi
  • Lebanese

Text to Speech (TTS) & Speech to Text (STT)

Clause-streaming Arabic voice synthesis across 8 preset voices, paired with dialect-tuned Arabic speech recognition and number reading.

  • 8 ready preset voices (male & female)
  • Clause-level voice streaming with low latency
  • Arabic speech recognition tuned for local numbers

Voice Cloning, Sentiment & Call Analysis

Consented voice cloning from a 3–6 second clip, plus post-call analysis covering caller sentiment, summaries, and structured fields.

  • Consented voice cloning only
  • Post-call sentiment analysis & conversation summary
  • Extracted fields and tool-call audit log per session

The selected dialect stays consistent for that agent, so a customer does not hear Jordanian and Gulf mixed in one reply.