Acme — a grounded support agent you can embed in your product (in minutes)

This is a forkable template for adding an AI support agent to your own product. "Acme" is a fictional project-management SaaS used as the example — swap the knowledge base for your own docs and you have a support agent grounded in your facts.

How it was built (write-up + video): <dev.to link> · Try it on your own product: syntheticbrew.ai

What you get

• An agent that answers questions only from your docs — and says "I don't have that in the docs" instead of hallucinating when it doesn't know.
• Per-visitor memory — it remembers what a user told it (team size, plan, needs) across sessions, isolated per visitor.
• A drop-in chat widget — one <script> tag on any page.

The point: you don't write provisioning code

You don't hand-wire API calls. You install the SyntheticBrew MCP in your coding agent (Cursor, Claude Code, etc.) and tell it — in plain English — to build the agent from the docs in this repo. The agent does the setup and hands you back the embed snippet.

That's the whole pitch: your coding agent provisions the support agent for you.

Deploy your own in ~5 minutes

1. Fork this repo and replace knowledge-base/*.md with your own product docs (plain markdown — pricing, limits, policies, FAQs; see the samples for the shape).

2. Install the SyntheticBrew MCP in your coding agent — one line, from the catalog listing or docs.

3. Paste the prompt in PROMPT.md to your coding agent. It creates the agent, uploads your docs as a knowledge base, links them, and returns a widget snippet.

   (New here? Hit "Use this template" above to get your own copy in one click.)

4. Drop the snippet on your site (see embed-example.html).

Free tier is enough to try the whole thing end to end.

What's in here

Path	What it is
knowledge-base/	7 sample docs = the agent's source of truth. Replace with yours.
PROMPT.md	The instruction you give your MCP-enabled coding agent.
embed-example.html	The widget <script> tag, with placeholders.

Honest notes

• Grounded Q&A (the bulk of support value) is solid — it looks up your docs before answering and stays honest when a fact isn't there.
• Memory is reliable when a user explicitly asks it to remember something or states a clear update. Don't market it as "it magically remembers everything" — passing mentions aren't always captured. Set the expectation accordingly.

Built on SyntheticBrew — an embeddable agent runtime.