Connect your agent

We never run your code. Your agent stays where it is; we send it messages over the internet and read its replies.

Only a test or staging agent

Before any test, you confirm in writing that the agent you connect is a test or staging agent, not one wired to live payments, real refunds, real emails or real customer records. If you change its address, you confirm again: a confirmed staging address can never be quietly swapped for a production one.

The four ways we can reach it

On your engagement page, choose Connect an agent and pick the format your agent already speaks:

1. Web API (Duelpoint's simple JSON format): we send the conversation so far; your agent replies with its next message.
2. OpenAI-compatible chat API: the chat-completions format many agents already accept.
3. MCP tool server: we call your agent's chat tool over the Model Context Protocol.
4. Agent-to-agent (A2A): we talk to your agent over the A2A protocol.

Give it a name, its public https:// test address, and an API key if it needs one. The key is encrypted the moment you save it, is never shown again, and is deleted when each test ends (Continuous testing: kept sealed for the months you agreed to).

The full technical details, with copy-paste examples in each format, are in the developer guide on the website (Developers > Connect your agent).

What your agent must do

• Be reachable at a public https:// address with a normal certificate.
• Answer within 30 seconds, with a reply under 1 MB.

We only ever call public addresses: never private, internal or cloud addresses, not even after a redirect.

Fake tools

If your agent takes actions (looks up orders, issues refunds, takes payments), point it at fake tools while it is being tested, so nothing real can happen. We host a fake order, refund and payment service for this; every answer it gives is marked as fake. On a confirmed test agent's card, press "Turn on our fake tools" to get its secret address. The tools answer only while one of our tests is waiting for your agent's reply. With them on, what they recorded is the truth about refunds and payments, and the report says so ("from your agent's tool records (exact)").

Who makes your agent's model

When you connect an agent you can say who makes the AI model behind it (Anthropic, DeepSeek, another company, or not sure). We judge honesty and tone with a model made by a different company, so a judge never favours its own maker. If you leave it as "not sure", the judge may share a maker with your agent; the report says so, and a person spot-checks an extra sample.

How many calls we make

Every test has a fixed maximum number of messages it may send to your agent, worked out from what you bought. The run stops cleanly if it ever reaches it, so a test can never run up your own costs.