Build the agent
Train your agent from ChatGPT or Claude
Connect the assistant you already use and train your agent by chatting. Tell it how you want shoppers handled; it reads your existing rules, writes new ones in our format, checks them, and saves them once you agree.
What it can do
- Set up an agent from your website address, tell you when it has finished reading the site, and give you the code to install it. On a connected Shopify store it gives you the app embed steps instead.
- See the agents you can edit and the training rules they already have.
- Turn plain instructions ("when someone buys the harness, suggest the leash") into rules, and check them for gaps, duplicates and clashes with your saved rules before saving.
- Rewrite an existing rule instead of adding a near copy of it.
- Switch rules off or delete them, after asking you first.
Saved rules go live straight away and appear under Knowledge sources → Training with an Added from assistant label. You can edit, switch off or delete them there like any other rule.
Connect
Use this server URL:
https://agentmatica.com/api/mcp/train- ChatGPT: in Settings, add a custom app (connector) with the URL above. On some plans this sits behind developer mode.
- Claude: in Settings → Connectors, add a custom connector with the URL above. Step by step: Claude connector.
- Claude Code, Cursor and other MCP clients: add it as a remote HTTP server; no API key is needed.
Your assistant opens an Agentmatica sign-in page. Sign in, check that the request came from the assistant you are connecting, and approve. Then ask it something like "show me my agent's rules" to get started.
No Agentmatica account yet? The sign-in page creates one for you on the Free plan, then brings you back to approve the connection.
Reading rules happens straight away. Before it saves, changes, switches off or deletes a rule, ChatGPT asks you to allow it. Click Allow once, or Always allow to skip the question next time. Until you click, nothing is saved, and the chat can look as if it has stopped.
Safeguards
- It works as you: it can only reach agents you can edit, and viewers stay read-only.
- Every rule goes through the same checks as rules saved in the dashboard, including the 100-rule limit. A rule that clashes with one you already have is refused unless you confirm it.
- It has no tools for chats, customers, orders, billing or settings. Agents it creates count toward your plan, the same as ones made in the dashboard.
Disconnect
Go to Account → Connected assistants and choose Disconnect. The assistant loses access straight away; rules it already saved stay until you remove them.