Test your AI agent with MCP

Managers can ask Claude or ChatGPT to try out an AI agent in its playground: see what it would reply, without anything reaching your customers.

Written by Anders Eiler
Last updated 2026-11-18
On this page

With the Herodesk MCP server, a manager can test an AI agent from their AI assistant, the same way as in the agent's Playground in Herodesk. Your assistant sends a customer message to the agent and shows you what the agent would reply. You need a connected assistant with write access, see What is the Herodesk MCP server?, and the manager or admin role.

What happens when you test

  • Nothing reaches your customers. No message is sent and nothing is saved on a conversation.
  • No resolution is billed for a test.
  • Lookups are real. When the agent looks up an order, a product or a booking, it asks your real integrations, exactly like it would on a live channel.
  • Actions are only proposed. If the agent wants to cancel an order, make a refund or change a booking, your assistant shows the proposed action. It is never carried out.

Four ways to test

ModeExample prompt
Free chat"Ask our AI agent Nova, as an email customer: Where is my order?"
As a test customer in live chat"Test Nova in live chat as the customer camilla@example.dk and ask where her order is."
Preview an existing conversation"What would Nova have replied in conversation #4821?"
Follow up on a preview"Continue that conversation: the customer now asks when the parcel arrives."

You can choose the channel to simulate: live chat, email, Messenger, Instagram or WhatsApp. The agent then writes like it would on that channel.

When you test as a customer in live chat, the agent treats that person as verified, so order and return lookups use that customer's real orders. These tests are recorded in the audit log.

A preview only works on conversations you can see in Herodesk. Conversations in inboxes outside your teams can't be previewed.

What your assistant sees

Besides the reply, your assistant can see whether the agent would hand the conversation over to your team and why, which knowledge it used, which topic it matched and the actions it proposed. Ask for it: "Which sources did Nova use for that answer?"

Limits

You can run 60 tests per hour for your whole workspace through MCP and the API. Tests in the Playground inside Herodesk don't count towards that limit. When the limit is reached, your assistant is told how long to wait.