# Creating and Managing an MCP Endpoint

An MCP Endpoint is the [Agent Channel](/products/ai-studio/agent-channels/) an MCP client connects through. It publishes one shared work folder as a Model Context Protocol server: a client such as Claude or another MCP-compatible agent can list the folder's Actions, run one, or hand Treasure AI Studio a task in plain language, and read back the result. The work runs in Studio, against the folder's files and with the executor's connections. The client only delegates the work and reads back the result.

This page covers creating an endpoint, storing the URL and header the client needs, and running the endpoint afterward. [Connecting an MCP Client](/products/ai-studio/agent-channels/mcp-clients) covers the client side: configuring it, the six tools it gets, and how to read their results.

Set one up when people already work in an MCP client and want to use your team's Actions from it without opening Studio. The endpoint is created once, the client is configured once, and the folder's Actions become tools the client can call.

## Prerequisites

- A [shared work folder](/products/ai-studio/shared-folders/) that you **manage**. The wizard offers only folders where you hold a work mode with manage rights, and an archived folder cannot be chosen.
- Permission under the account's creation policy. SDK Apps and MCP Endpoints share one policy, shown on the Agent Channels page under **Who can create SDK Apps and MCP Endpoints** — see [Choose Who Can Create Agent Channels](/products/ai-studio/agent-channels/sdk-apps#choose-who-can-create-agent-channels). When it is **Account admins only** and you are not one, **New MCP Endpoint** is disabled.
- Treasure AI Studio on **Web or Desktop**, where Agent Channels are created and managed. They cannot be set up from the Mobile app.
- An MCP client that can connect to a remote server over Streamable HTTP and send a fixed `Authorization` header. The token is shown once, so have the client's settings or a secret manager ready.


Decide who the executor should be before you start. The creator becomes the executor, and every Action and chat the endpoint runs uses that person's credentials, permissions, and credits. Anyone who holds the endpoint's token gets that reach over the folder, which is why [what a channel chat is not restricted from](/products/ai-studio/agent-channels/channel-chats#what-is-not-restricted) is worth reading first.

## Create an MCP Endpoint

1. Open **Settings → Agent Channels**. MCP Endpoints have their own section, **MCP Endpoints**, below **SDK Apps**. Before the first endpoint is created, the section shows a **Connect your MCP client** guide in place of the list.


MCP Endpoints section before any endpoint exists, with the New MCP Endpoint button and a three-step Connect your MCP client guide
1. Select **New MCP Endpoint**. The wizard opens at step 1 of 2 and first checks which shared work folders you can use. If you manage none, it says *"Creating an SDK App or a static-token MCP endpoint requires manage access to a shared work folder. Ask a work folder manager to grant you manage access."* instead of showing the form.
2. Under **Endpoint name**, enter a name — for example, *"Marketing team MCP"*. Name it after the people or the client that will use it, because the name is how the endpoint is told apart in the list. Up to 255 characters.
3. Optionally add a **Comment** for your colleagues, such as the owning team and how to reach them. Up to 1,000 characters. Disabling the endpoint or stepping down as executor can overwrite it with a note, so treat it as an operational note rather than a permanent record.


Step 1 of 2 of the New MCP Endpoint wizard with an endpoint name entered and the optional Comment field
1. Select **Next**, then choose the **Shared work folder**. The field searches as you type, and only folders you manage are listed.
2. Read the consent text that appears under the folder. It states what you are authorizing: anyone holding a token can use that folder and its Actions, the folder is read-only, and the chats run with your credentials, permissions, and cost in a temporary work area that is discarded when each chat ends. **Executor** is shown as you and cannot be set to someone else here.
3. Choose an **Expiration** for the first token: **7 days**, **30 days**, **90 days** (the default), a custom date, or **No expiration**. A custom date resolves to the end of that day in the time zone set on your Settings page. The expiration is fixed when the token is issued — see [Issuing a Gateway Token](/products/ai-studio/agent-channels/tokens#issuing-a-gateway-token).


Step 2 of 2 of the New MCP Endpoint wizard with a shared work folder selected, the executor shown as the current user, the consent text for that folder, and an Expiration of 90 days
1. Select **Create endpoint & issue first token**. The endpoint is created, then its first token is issued, and the completion screen appears.


An MCP Endpoint has no advanced settings. Browser access and Chat Access Token lifetimes belong to SDK Apps and do not apply here: an MCP client authenticates every request with the endpoint's token directly, and the gateway always refuses calls from a browser.

If the endpoint is created but the token is not
The two actions are separate requests, so the second can fail on its own. When it does, the wizard shows **Endpoint created — token pending** rather than starting over: the endpoint exists. Choose **Retry issuing a token**, or **Issue a token later from the endpoint page** and issue it from the endpoint's detail page when you are ready.

## Store the Connection Values

The completion screen is headed **Connect your MCP client** and shows the two things the client needs:

| Field  | Value  |
|  --- | --- |
| **Connection URL** | `https://<your account's Agent Gateway>/v1/mcp/<endpoint ID>`. The URL is issued by Treasure AI Studio, so copy it rather than composing it yourself |
| **Authentication header** | `Authorization: Bearer tasgwt_<the secret>`. **Copy** copies the whole line; **Copy value only** copies just `Bearer tasgwt_<the secret>`, for clients that ask for the header name and value in separate fields |


Connect your MCP client completion screen with the Connection URL and its Copy button, the Authentication header line Authorization: Bearer with Copy value only and Copy buttons, a warning that the token will not be shown again, the I have stored the values checkbox, and a disabled Done button
This is the only time the token is shown
The secret is never displayed again, and Treasure AI Studio cannot recover it. Put the authentication header into your MCP client's settings or a secret manager before you leave the page. If you lose it, revoke the token and issue a new one.

1. Copy the URL and the header into your MCP client — [Connecting an MCP Client](/products/ai-studio/agent-channels/mcp-clients#configure-the-client) shows where they go.
2. Select **I have stored the values**. The **Done** button stays disabled until you do.
3. Select **Done**. The endpoint's detail page opens.


Leaving the page before checking the box prompts *"You haven't saved this token yet"* and offers **Go back**. Take it unless you really have stored the token.

The URL does not change while the endpoint exists, and the detail page shows it again as **Connection URL**. The token in the header is the only part you cannot see again.

## Read the MCP Endpoints List

Each endpoint is one row under **MCP Endpoints**, with the same columns as SDK Apps:

| Column  | What It Tells You  |
|  --- | --- |
| **Name** | The endpoint's name, with its comment underneath when one is set |
| **Folder** | The single shared work folder the endpoint publishes |
| **Executor** | The user whose credentials, permissions, and credits the endpoint runs with, shown by email. **No executor** in amber means no one holds the role and the endpoint cannot run |
| **Status** | The endpoint's current state (below). A disabled endpoint adds *"Connections are stopped"* |
| **Eye icon** (last column, unlabelled) | Opens the endpoint's detail page |


An account can have several MCP Endpoints, including more than one on the same folder. Each has its own URL and its own tokens, and a token for one endpoint is refused by every other — so giving each team or client its own endpoint lets you stop one without touching the rest.

MCP Endpoints list with two endpoints on the same shared work folder: one Active and one Stopped — needs an executor with No executor in amber
Every member of the account sees every row. Unless you manage the folder, are the executor, or are an account administrator, the **Folder** and **Executor** cells show only a dash and a lock icon. On the detail page those members see the name, comment, status, and **Connection URL**, with a notice that the folder, executor, and tokens are visible only to members with access. The URL is not a secret: without a token, it gets nothing but a refusal.

MCP endpoint detail page as seen by a member without access: the name, the Active status, the Connection URL with a Copy button, and a notice that the shared work folder, executor, and tokens are visible only to members with access
## Endpoint Statuses

| Status  | Meaning  | How to Clear It  |
|  --- | --- | --- |
| **Active** | The endpoint accepts connections | — |
| **Disabled** | Someone switched the endpoint off. Every request is refused; no token is revoked | Turn the switch back on |
| **Stopped — needs an executor** | The executor stepped down, so the seat is empty. Nothing runs, and the endpoint cannot be enabled while the seat is vacant. Tokens are retained | A manager of the folder selects **Become executor**, then enables the endpoint |


An endpoint also refuses connections, without changing status, if the executor loses their work mode on the folder. The check runs on every request, not only when the endpoint is saved. When a manager deletes the folder itself, its endpoints are deleted with it — see [Delete an Endpoint](#delete-an-endpoint).

## Manage an Endpoint

Open an endpoint from the list. What you can do depends on who you are:

| Role  | What They Can Do  |
|  --- | --- |
| **Executor** | Edit, issue and revoke tokens, disable and enable, delete, and step down |
| **Manager of the endpoint's folder** | Edit, issue and revoke tokens, disable and enable, delete, and become the executor |
| **Account administrator** | Disable and enable only. Being an administrator alone does not allow editing, deleting, or issuing and revoking tokens |


The detail page shows the endpoint's **Overview** — shared work folder, executor, comment, and the consent it was created on — and lists the tools the client gets: `list_work_folders`, `list_actions`, `run_action`, `start_chat`, `continue_chat`, and `get_chat_result`. The list is the same for every endpoint and cannot be narrowed.

Detail page of a disabled MCP endpoint: the Enabled switch turned off, the Connection URL, a notice that connections are stopped, the Overview with shared work folder, executor, comment, consent text and available tools, the Tokens table with Issue token, Resign as executor, and Delete endpoint in the danger zone
### Disable and Enable

Use the **Enabled** switch in the header to stop an endpoint without destroying anything. Turning it off opens **Disable MCP endpoint**, which asks for an optional comment. Write down why, because that note is how the next person decides whether it is safe to turn back on. Requests stop immediately; tokens are not revoked. Turning the switch back on takes effect at once, without a dialog, and clients can use any token that has not expired or been revoked.

Disabling is the right tool for an incident: it is immediate and reversible, and it is the one action an account administrator can take on an endpoint.

Disable MCP endpoint dialog explaining that connections stop immediately and tokens are not revoked, with an optional comment field and a Disable button
### Edit the Name or Comment

**Edit** in the **Overview** changes the endpoint's name and comment. Nothing else can be edited. The folder and the executor are what the endpoint was consented to on, so a different folder means a different endpoint, and a different executor means a handover (below).

Edit MCP endpoint dialog with Name and optional Comment fields and a Save button
### Take Over or Step Down as Executor

The executor role carries the consent, so it changes hands only when each person acts for themselves.

- **To step down**, select **Resign as executor**. The seat empties and connections stop in the same action; existing tokens are kept but cannot be used until someone takes the role and enables the endpoint. Add a comment so the folder's managers know what happened.
- **To take over**, open the endpoint and select **Become executor**. You must manage the endpoint's folder, and you can take over whether or not the seat is currently empty. The dialog warns that future calls through existing tokens will use your credentials, permissions, and cost. Taking the role does not restart a stopped endpoint — enable it afterward, when you are ready for that.


Resign as executor dialog explaining that the executor seat is vacated and connections stop, with existing tokens kept until a work folder manager becomes executor and enables the endpoint, and an optional comment field
Detail page of an MCP endpoint with the status Stopped — needs an executor: the switch off, the Overview showing No executor, a notice that a work folder manager must become executor and enable the endpoint to resume connections, and a Become executor button
There is no way to appoint someone else. Handover is a claim, not an assignment.

### Issue and Revoke Tokens

The **Tokens** section lists every token issued for the endpoint, active and inactive together, with its hint, status, issue date and issuer, expiration, and last use. **Issue token** opens the same expiration choices as the wizard and then shows the **Connect your MCP client** values again, for the new token only.

Revoking a token stops new calls with it immediately and cannot be undone. An endpoint can hold several tokens at once, which is how you rotate one without downtime — see [Rotating a Gateway Token](/products/ai-studio/agent-channels/tokens#rotating-a-gateway-token).

Revoke connection token dialog naming the token hint, explaining that new calls with it stop immediately and that revoked tokens cannot be reused
### Delete an Endpoint

**Delete endpoint**, in the danger zone at the bottom of the detail page, permanently removes the endpoint. Its URL and every token under it stop working. You are asked to type the endpoint's name to confirm. When a manager deletes a shared work folder, every endpoint on it is deleted the same way.

Deleting the endpoint does not delete the chats it created; they stay in the executor's chat list under the account's normal retention. Deleting the shared work folder is different: it deletes the folder's endpoints and the chats they created. Prefer disabling when you might want the endpoint back. Deletion cannot be undone, and a new endpoint has a new URL that every client has to be reconfigured with.

## Troubleshooting

| Symptom  | Cause and Fix  |
|  --- | --- |
| The wizard shows *"Creating an SDK App or a static-token MCP endpoint requires manage access to a shared work folder"* instead of the form | You hold no work mode with manage rights on any shared work folder. Ask a manager of the folder you need to grant you one, or to create the endpoint |
| **New MCP Endpoint** is disabled, or the wizard says only account admins can create them | The creation policy is set to **Account admins only** and you are not an account administrator. Ask an account administrator to create the endpoint or to change the policy |
| The endpoint shows **Stopped — needs an executor**, or the **Enabled** switch will not turn on | The executor seat is empty. A manager of the folder selects **Become executor**, then enables the endpoint |
| The client is refused with `401` on every request | The token is revoked or expired, it belongs to a different endpoint, or the endpoint has been deleted. Check the token row on the detail page, and that the URL and token came from the same endpoint |
| The client is refused with `403` | The endpoint is disabled or has no executor, or the executor has lost access to the folder. Check the endpoint's status. A `403` also comes back when the request carries an `Origin` header — see [Connecting an MCP Client](/products/ai-studio/agent-channels/mcp-clients#troubleshooting) |
| You lost the token | It cannot be recovered. Issue a new one, update the client, and revoke the old one |


## Next Steps

- [Connecting an MCP Client](/products/ai-studio/agent-channels/mcp-clients) — configure the client, and the tools and results it works with.
- [Gateway Tokens and Chat Access Tokens](/products/ai-studio/agent-channels/tokens) — expiration, rotation, and revocation.
- [What an Agent Channel Chat Can Reach](/products/ai-studio/agent-channels/channel-chats) — what the endpoint's chats can and cannot do.
- [Actions](/products/ai-studio/skills/actions) — create the Actions the endpoint publishes.