Oro AI Agent

Oro is an AI agent for small businesses, live in production. It connects a shop's Messenger, Instagram and Telegram, answers customers from the shop's own catalogue and rules, and hands the conversation to the owner when it should. Small shops sell through Messenger, and their customers rarely type a full question. They send a voice note on the way to work, or a photo of something they saw and ask "do you have this?". Most chat assistants drop the voice note and only guess at the photo. During the hackathon we built Oro on Muse: 1. Muse Spark 1.3 answers the customers. Owners pick it in the Oro app as "Oro Muse", a reply tier in the agent's Model box. It runs through Oro's reply engine, with tool calls over the shop's catalogue. On our bench of 24 scripted customer turns in English, Burmese and Thai, Muse Spark 1.3 passed all 24, the same as our previous model, at a median of 3.7 s per turn. 2. Voice notes are heard. A customer's voice note is transcribed by muse-voice-transcribe-1.0 before the assistant answers. Messenger sends MP4/AAC and the endpoint takes only 16 kHz mono WAV, so we built a 2 MB audio-only ffmpeg into our image (Debian's package would have added 429 MB) and convert through files, never pipes. 3. Photos are recognised. SAM 3.1 finds the item, using the shop's own kinds of things as its phrases. Muse Spark describes what is inside the box as search words. And Muse Spark is shown the customer's photo beside the shop's own product photos, up to 40 in one call, to say which one is the same piece, because a catalogue's words rarely say what a piece looks like. The assistant names the match and sends its picture. 4. A public, dependency-free TypeScript client for these models, and a Messenger bot anyone can run with a Page token and a Meta key. It holds up in production: Oro Muse is a tier an owner chooses, one switch withdraws it in seconds with no deploy, a voice note or photo that cannot be read never drops the customer's turn, and if a Meta call fails Oro answers that call with its own model, so a customer always gets a reply.