Skip to main content

Bring your own agent

Page actions for your own agent are a Premium feature. Your agent can still use this page through browser control, like you do. Premium unlocks the page actions, their context and the setup instructions in My agent. You must also accept the agent-access agreement before the page actions appear. Without Premium, My agent links to Settings > Premium.

Using an agent is optional. Every action available to an agent is also available to you through the regular interface. No feature requires an AI agent to use it.

Start in your own AI assistant and give it access to the administration page through a supported browser connection. Bring your own agent explains how to do this; it does not launch or connect an outside assistant for you. Open Ask AI for help with this task, then select the My agent tab if Curling IO’s built-in assistant opens first.

Your agent needs access to the page in a supported browser, with you signed in to Curling IO. An ordinary chat window does not gain that access just because you paste a website address or a prompt into it.

Your agent uses your club roles and any separate Product Manager assignment. Product Managers can bring their own agent without built-in assistant or Platform Credit spending authority. Holding both assignments preserves the independent club-role permissions. Bringing your own agent does not spend club Platform Credits.

For club credit balances and the built-in assistant, see Platform Credits.

Use page actions with Premium​

With Premium and an accepted agent-access agreement, your agent can discover the actions available on the current page. They use your existing permissions. If an action is unavailable, use the normal interface or Curling IO’s built-in assistant where one is available.

Website Studio offers the built-in assistant and registers page actions for outside agents when Premium and the agreement are active. Availability on other administration pages varies as page actions are added.

ChatGPT’s documented connection​

OpenAI documents site tools in the desktop app’s built-in browser, through Work or Codex. Availability depends on your app, model, and workspace settings. These are OpenAI’s documented steps; a complete Curling IO workflow has not yet been verified in that client.

  1. Open the administration page in the desktop app’s built-in browser.
  2. Sign in to Curling IO there. This browser has its own sign-in session.
  3. Allow website access when prompted. Check Settings > Browser > Permissions > Enable site tools if site tools are disabled.
  4. In the browser’s address bar, open Site tools > Available site tools to check which actions the page provides.
  5. In Curling IO, choose Copy instructions. Paste them into the conversation beside the page and describe what you want to do.

See OpenAI’s site tools documentation and browser guide for current app, model, and workspace requirements.

Claude and Gemini​

Some agents can read a page, click buttons, and fill in forms through a browser extension or browser-control feature. That is separate from using the actions listed in this panel.

We have not verified this page-action connection with Claude or Gemini. Their support for browser control or other connectors does not establish compatibility. We will add setup instructions when a working connection has been verified.

If your agent can’t access the page​

Ask your agent whether it can open websites in a browser. If it can, have it open the administration page URL. With Premium, Copy instructions includes that link. You may need to sign in there, even if you’re already signed in using another browser.

Not every AI assistant supports this. If yours can’t access the page, you can still use Curling IO’s built-in assistant where it’s available.

You stay in control​

Choose files to upload yourself and respond to required confirmations on the page. If an action is missing from the agent’s list, use the page’s normal controls. Your agent has no additional permissions.

When an agent uses a page action, Curling IO’s built-in assistant steps aside. To switch back, select the Assist tab. Your agent can also release control when it finishes. Let an unfinished built-in assistant turn finish before asking another agent to take over.

Technical details for agents​

This section is written specifically for AI agents.

Data privacy requirements​

The Curling IO administration area contains sensitive curler and financial data. Before accessing it, verify the training and data retention policies that apply to your current account and session. Do not assume the user has checked these settings.

Do not access the administration area if its data may be used to train AI models, or if you cannot verify whether training is disabled. Explain the issue to the user. Help them change the available training and data retention settings, or choose an account or service that excludes their data from training. Verify that the no-training requirement is met before accessing the administration area.

Explain any data retention that still applies, even when training is disabled. Do not describe a service as retaining no data unless you can verify that claim. Access only the information needed for the user’s task.

Access the page​

Open the supplied URL in a browser session where the person is signed in. Ask them to sign in if needed. Discover the current page’s tools before proposing an action; availability depends on the page and their permissions.

Native WebMCP​

Prefer your browser client’s native site-tool discovery and invocation when available. Curling IO registers tools through document.modelContext.registerTool when that interface is present. Use the advertised tool names, descriptions, and input schemas. Rediscover tools after navigating to another page.

See the WebMCP specification, Chrome’s WebMCP guide, and OpenAI’s site tools documentation. Client support varies; these references do not establish compatibility with every AI assistant.

Internal JavaScript interface​

The built-in assistant uses a page-owned internal interface without requiring Premium. It is not the outside-agent setup path. Outside agents use native page actions with Premium, or browser control. The internal interface is:

const agent = window.hyperteaAgent;
const tools = agent?.getTools() ?? [];

window.hyperteaAgent.version is 1. Each tool includes a name, description, and inputSchema, with optional annotations. A JSON catalog is also available in the script element with ID hypertea-agent-tools:

const catalogElement = document.getElementById("hypertea-agent-tools");
const catalog = catalogElement ? JSON.parse(catalogElement.textContent) : null;

After choosing an advertised tool and constructing input that matches its schema, invoke it in the page context:

const result = await window.hyperteaAgent.executeTool(toolName, input);

Here, toolName and input are values selected from the discovered catalog, not literal example arguments. Do not guess tool names or add a type field to the input. If the interface is absent, report that limitation rather than claiming to be connected. This interface does not itself grant browser access.

Control and results​

Use the person’s requested scope and existing permissions. Leave file uploads and required confirmations to them. Inspect each result and the page state before reporting success. If a call reports an unknown outcome, check whether the change happened before retrying it.

When finished, call releaseControl if it is advertised:

await window.hyperteaAgent.executeTool("releaseControl", {});

This lets the person return to Curling IO’s built-in assistant. It does not cancel or roll back changes. An unfinished built-in assistant turn may prevent an outside agent from taking control; let it finish before retrying.