Test Your AI Agent from Claude Code or Cursor — Without Contacting a Single Customer

The short version

The test_chat tool on the ThinnestAI MCP server sends your agent a message as a website visitor and returns its reply — plus every tool it used, with what went in and what came out. It is the real agent: the same instructions, model, knowledge and actions your customers get.

What it cannot do in a test is reach a person. Tools that would phone, send a WhatsApp, email, send a code or change a calendar booking are switched off.

Why the test has to be the real thing

A test that runs a copy of your agent tells you how the copy behaves. When you change an instruction, add knowledge or connect an action, the question is whether the agent your customers talk to now gets it right. So test_chat goes through the same path as the chat widget on your website — and what you test is what ships.

What is switched off, and what is not

Switched off in a test:

• Calling someone now, or scheduling a callback
• Sending a WhatsApp, an email, a form or a verification code
• Holding, booking, rescheduling or cancelling a calendar meeting

Still runs for real:

• Your knowledge
• Your own actions — your API is really called
• Reading calendar availability
• Lead capture, and escalation to your own team

When the agent would have used a switched-off tool, it says so plainly — "I would schedule a callback for +91… at 5pm" — and carries on. You see the behaviour; nobody's phone rings.

Your own actions run for real on purpose: checking that the agent calls your order-lookup endpoint with the right order number is the point of the test. If an action of yours changes data, test against a staging endpoint.

A loop that stays in the editor

1. "Add our refund policy page to the support agent's knowledge."
2. "Ask it: can I return an opened item after 10 days?"
3. The assistant reads the reply and the tools used, and spots the agent answered from the wrong page.
4. "Tighten the instructions so it quotes the refund policy, and test again."

Pass back the session from the last answer to continue the same conversation — useful for testing a follow-up question or a hand-over — or leave it out to start fresh.

Good to know

• Test chats appear in your inbox, like Playground chats, so your team can see them.
• They use your workspace's message allowance, like any other chat.
• A Build key can test-chat — it is the key for building, and testing is part of building.

Test your agent from your editor →