# Connecting Claude to Treasure AI Studio

You can add Treasure AI Studio to Claude on the web (claude.ai) as a custom connector, so you can run Studio agents and Actions from a Claude chat. Each request runs with the work folders, permissions, and cost of the person who connected. This page covers the setup in Treasure AI Studio and in Claude, and what happens when you send a request. Use it when you already work in Claude and want audience information or the output of a Studio Action there, without switching tools.

## How the Connection Is Set Up

Two people set up the connection, in two places.

| Who  | Where  | What They Do  |
|  --- | --- | --- |
| **Account administrator** | Treasure AI Studio, under **Settings → Agent Channels** | Creates one OAuth endpoint and copies its Connection URL |
| **Each person who uses Claude** | Claude, under **Settings → Customize → Connectors** | Adds the Connection URL to Claude, signs in to Studio, and allows access |


Only account administrators can create OAuth endpoints. Everyone else connects to an endpoint that an administrator has already created. Each person also needs a Treasure AI Studio account in the same account, with access to the work folders they want to use. In Claude, each person needs an account that shows **Add custom connector** under **Customize → Connectors**. If it doesn't appear, check your Claude plan and your organization's settings in Claude.

## Step 1: Create an OAuth Endpoint in Studio

An account administrator creates one OAuth endpoint, and everyone who connects uses its Connection URL. For the full setup, see [Create an OAuth Endpoint](/products/ai-studio/agent-channels/mcp-oauth#create-an-oauth-endpoint).

1. Open **Settings → Agent Channels** and select **New MCP Endpoint** in the **MCP Endpoints** section.
2. Under **Connection method**, select **OAuth · connected user**.
3. Enter an **Endpoint name**, such as *Studio MCP OAuth*, and add an optional **Comment**.
4. Select **Create endpoint**.
5. On the **OAuth endpoint created** screen, select **Copy** next to the **Connection URL**.


The Connection URL has the form `https://<your account's Agent Gateway>/v1/mcp/<endpoint ID>`, for example `https://agent-gateway.us01.treasure.ai/v1/mcp/<endpoint ID>` on the US01 site. The endpoint's detail page shows its status as **Active**. Give the Connection URL to the people who will set up the connection. This guide uses the endpoint's own Connection URL. The common MCP URL is an alternative, described in [The Common MCP URL](/products/ai-studio/agent-channels/mcp-oauth#the-common-mcp-url).

## Step 2: Add Treasure AI Studio to Claude

Each person does this in Claude.

1. In Claude, open the account menu at the bottom left and select **Settings**.
2. Select **Customize**, select **Connectors**, and select **Yours**.
3. Select **+ Add**, and then select **Add custom connector**.


1. Enter a name, such as *Treasure AI CDP*, and paste the Connection URL from Step 1. Select **Continue**.


1. Keep the settings that Claude detects: **Authentication** set to **Sign in now**, and **OAuth client** set to **Register automatically**. Leave **Request headers** empty. Do not add a token. Select **Add**.


1. The connector's page says that you're not connected yet. Select **Connect**. Your browser opens Treasure AI Studio.


1. If Studio asks you to sign in, sign in. On **Authorize MCP access**, check that the callback host shown at the top is `claude.ai` and the application name is **Claude**. Check the requested MCP resource, which is the Connection URL from Step 1, and the account. Select **Allow access**.


The requested permissions are to view work folders, list Actions, read execution results, run Actions, and start or continue chats. These are the same permissions described in [The MCP Permission](/products/ai-studio/agent-channels/mcp-oauth#the-mcp-permission).

1. Check **Customize → Connectors → Yours**. The connector appears with type **Web**, a **Custom** label, and a checkmark in the **Status** column.


## Step 3: Review the Tool Permissions

Open the connector from **Customize → Connectors → Yours**. Under **Tool permissions**, Claude lists the six Studio tools in two groups: three **Read-only tools** and three **Write/delete tools**. Each group is set to **Needs approval**, so Claude asks you before it uses a tool. Claude explains on this screen that in **Auto** mode, an automated check can let Claude use tools without asking, even tools set to **Needs approval**. To approve each request yourself, use **Manual** mode.

## Step 4: Send a Request from Claude

1. Start a new chat. Select **+** beside the prompt, select **Connectors**, and make sure the toggle for your connector is on.


1. Send a request that needs Studio, such as *Use Treasure AI CDP to list the Studio work folders I can access. Do not run Actions or change anything.*
2. When Claude asks to use a tool, review the tool name and description, and choose **Decline**, **Always allow**, or **Allow once**. Before you choose **Always allow**, check what the Action does in Studio.


If Claude answers without using a tool, check that the connector is turned on for the chat, and ask for something specific that needs Studio. A general explanation of what the connector can do does not show that Claude reached Studio.

## What Runs in Studio

Each request runs as the person who connected. It uses that person's work folders, permissions, and cost. Studio does not ask for approval before each request. The confirmation you see in this flow is Claude's prompt in Step 4, not a Studio prompt.

Requests that start or continue chats appear in your Studio chat list as ordinary chats, with the name of the work folder they ran in. They can change files in your work folders as far as your work mode allows. [The Chats Are the Person's Own Studio Chats](/products/ai-studio/agent-channels/mcp-oauth#the-chats-are-the-persons-own-studio-chats) describes this in more detail. `run_action` runs a saved Studio Action, so check its purpose and expected effects before you allow it.

## Stop the Connection

You can stop the connection for one person or for everyone. Stopping it in Studio refuses requests on the Studio side, and disconnecting in Claude stops Claude from using the connection.

- **Stop everyone (Studio).** Turn off the **Enabled** switch on the endpoint. Claude's requests are refused until you turn it back on.
- **Stop one person (Studio).** Remove the person from the account, or remove their access to the work folders involved.
- **Disconnect (Claude).** The person opens the connector under **Customize → Connectors → Yours** and selects **Disconnect**.


If Claude asks you to sign in again, complete the Studio consent for the intended account and endpoint. If authorization keeps failing, disconnect the connector and select **Connect** to authorize it again.

## Troubleshooting

| Symptom  | Cause and Fix  |
|  --- | --- |
| The consent screen says *"No accessible account has an OAuth endpoint for the requested MCP resource."* | The URL isn't an OAuth endpoint's Connection URL, or your accounts can't be offered for this endpoint. Create or select an endpoint whose connection method is **OAuth · connected user**, and paste its Connection URL into Claude. Otherwise, ask an account administrator |
| The connector is added but says you're not connected yet | Adding the connector and authorizing it are separate steps. Open the connector, select **Connect**, and complete the Studio consent |
| A tool set to **Needs approval** runs without asking | Claude is in **Auto** mode. Switch to **Manual** mode to approve each request |
| The consent screen shows an unexpected account or resource | Select **Deny**, check the account and the Connection URL, and connect again |
| Requests are refused with `403` and `mcp_disabled` | The endpoint is disabled. An account administrator enables it |
| Other errors after you connect | See the troubleshooting table in [OAuth MCP Endpoints](/products/ai-studio/agent-channels/mcp-oauth#troubleshooting) |


## Next Steps

- [OAuth MCP Endpoints](/products/ai-studio/agent-channels/mcp-oauth): permissions, the common MCP URL, and how access ends.
- [Connecting an MCP Client](/products/ai-studio/agent-channels/mcp-clients): the six tools and how to read their results.
- [What an Agent Channel Chat Can Reach](/products/ai-studio/agent-channels/channel-chats): what these chats can and cannot change.
- [Headless](/products/headless): the overview of connecting other AI tools to Treasure AI Studio.
- Other Headless guides: [ChatGPT](/products/headless/chatgpt), [Gemini](/products/headless/gemini), [Databricks Genie One](/products/headless/databricks-genie-one).