Agent Beans website

The homepage and live demo of Agent Beans, an always-on Meta ads growth team for small businesses and indie developers, built on Meta's Muse models for the Meta Global AI Developer Hackathon.

Site: pallas-ads.github.io/agent-beans

This repository holds only the website. The Agent Beans code lives in its own repository.

What is here

Path	Contents
site/	The homepage (a single HTML file), the script that builds the site, and the Google Apps Script that records demo sign-ups in a Google Sheet
web/	The Agent Beans web app, built in replay mode: it reads a recorded run and changes nothing
replay/northpaw/	A recorded live run for Northpaw Supply, an online pet store already running ads
replay/beans-active/	A recorded live run for Beans' Active, a new fitness app with no ad history
.github/workflows/pages.yml	Builds the site and publishes it to GitHub Pages on every push to main

The homepage embeds each replay at demo/<run>/. Because it is the real app, a visitor can click through the Team Room, deliverables, dashboard, ads, memory and harness of each run.

Build it locally

Requirements: Node 18 or later.

cd web && npm ci && cd ..
bash site/build.sh
python3 -m http.server 8000 --directory _site

Then open http://127.0.0.1:8000.

Update a replay

A replay is recorded with pgt replay from a running Agent Beans server, in the code repository. Copy the data/ and media/ folders it writes into replay/<run>/. When the app itself has changed, copy the code repository's web/ folder here as well, so the replay build reads the files the new export writes.

Data and license

Everything in the demo is synthetic sandbox data. The site calls no model and holds no key or real ad account ID, and pgt replay refuses to export anything that is not sandbox data.

The code in this repository is MIT licensed.

Built with Muse Spark Open. The recorded runs were made with Muse Spark Open, used under the Muse Model License, which asks for this notice.