# MCP Actions

> Give your AI Agent tools from an external MCP server. Connect one from the catalog or by URL, and choose exactly which tools it may use.

MCP (Model Context Protocol) is an open standard that lets an AI use another service's tools directly. Instead of you describing an endpoint and its parameters by hand, an MCP server publishes its own tools, and your AI Agent discovers and uses them.

Connecting one takes a single click for most services. Go to **Actions & MCPs**, click **+ Add Action**, and choose **MCP**.

## Connect from the catalog

The MCP screen opens on a searchable catalog of well-known services, grouped by category (**Popular**, **Customer support**, **CRM & sales**, **Scheduling**, **E-commerce**, **Payments & billing**, **Marketing & email**, **Communication**, **Productivity**, **Data & analytics**, **Search & web**, **Travel & booking**, **Design**, **Developer & infra**).

1. Find the service you want, by browsing or with the search box.
2. Click its tile. Quickchat checks the server and discovers the tools it offers.
3. If the service requires you to sign in, an **Authorization required** panel appears. Click **Connect with \<service\>**, approve access in the popup that opens, and the panel switches to connected.
4. The action is created, enabled, and its tools are listed. Your AI Agent can use them from the next message on.

:::note
After you start authorization, Quickchat keeps the action and the pending authorization if you close the Actions sheet or authorization popup. Reopen the action and click **Open authorization again** to continue, or **Cancel** to stop. If the provider finishes authorization after you close the sheet, Quickchat completes the connection in the background and enables the action when your plan has capacity.
:::

:::tip
Your browser must allow popups for the Quickchat dashboard, otherwise the authorization window cannot open. If it is blocked you will see a message asking you to allow popups and try again.
:::

## Connect a server that is not in the catalog

Click **Enter a URL manually** below the catalog, then:

1. Paste the **MCP server URL** (must start with `https://`).
2. If the server authenticates with a static token rather than a sign-in flow, open **Connection headers (optional)** and add the header it expects, for example `Authorization: Bearer <your-token>`.
3. Click **Connect**.

Use **Browse the catalog** to go back to the tile view. If a service you need is missing, **Missing an integration? Suggest one** opens a pre-filled email to our team.

## Choose which tools the AI can use

Once connected, the action lists every tool the server exposes, each with its own toggle. Turn off anything you do not want the AI to call — a booking server's read-only lookups can stay on while its cancellation tool stays off.

**Default tool activation** decides what happens when the server adds a *new* tool later: on, and new tools are available immediately; off, and you enable each one deliberately.

Some catalog entries carry a **Tuned by Quickchat** marker. Quickchat has tested these servers end to end and supplies a focused tool selection, the memory rules those tools need, and a **Thinking message** for each. Select **Use recommended** to see what the tested setup would change: the **Apply the recommended setup?** dialog lists the tools it turns on and off and the memory rules it sets, and warns if it would overwrite rules you edited yourself. Nothing is saved until you save the Action, and you can still adjust every tool afterward.

![The "Apply the recommended setup?" dialog: the tools it turns on and the memory rules it sets](https://docs.quickchat.ai/_astro/recommended-setup-dialog.EFLkzzpY.png)

:::tip
The **MCP action name** and **MCP action description** are labels for you and your team — they identify the action in the Actions list, and the AI never reads them. What steers the AI is the tool descriptions the MCP server itself publishes, together with your Agent's [Main Prompt](/ai-agent/identity/#main-prompt). To steer it, name the situations where the Agent should use this server's tools ("answer every project question with the Linear tools") and pin any values the model would otherwise have to guess ("always search the `acme-eng` workspace"). When you connect the same server twice, name each connection for its job ("Notion help center", "Notion support requests") rather than for the vendor.
:::

## Save an MCP tool result to memory

Expand a tool and use **Save to memory** when a value from its result will be needed later in the conversation. Add a JSONPath expression for the value and a **Memory key** to identify it. You can save up to 10 values from one tool.

![Save to memory on an MCP tool: two JSONPath rows saving values from the tool result](https://docs.quickchat.ai/_astro/mcp-save-to-memory.DZ99IF2d.png)

Memory keys are shared with the Agent's API Actions. If an API Action already saves to the same key, the **Memory key** field reports the clash when you save, so pick a different key.

The saved value becomes part of the conversation data the AI can read on later turns. For example, a booking tool can save an event URI so a cancellation tool knows exactly which meeting to cancel. MCP tools do not use the `{{metadata_<key>}}` placeholder shown for API Actions because the AI supplies their arguments directly.

Bulk **Enable all** and **Disable all** changes keep memory rules intact. **Reconnect** also keeps your tool choices and memory rules while refreshing the server's tool list.

## What visitors see

While a tool runs, the chat widget shows a short status line describing what the AI is doing. Each tool has a **Thinking message** field in the Tools panel: leave it empty and the line is derived from the tool's name (a tool called `search_flights` shows "Searching flights…"), or write your own ("Searching the help center"). Tuned servers arrive with these filled in for the tools they recommend. The line gives the visitor something meaningful during the pause instead of a bare typing indicator.

![The Thinking message field on an expanded tool, holding the wording the visitor sees](https://docs.quickchat.ai/_astro/mcp-thinking-message.gcGBd5Cp.png)

## Reconnect and disconnect

- **Reconnect** — runs the authorization again. Use it if the connection expired, if you want to authorize a different account, or if the server's tool list changed and you want to pick the new tools up.
- **Disconnect** — revokes the stored authorization and takes the action offline. The action itself stays in your list, so you can reconnect later without setting it up again.

Changing the **MCP server URL** on a connected action clears its authorization and switches the action off, because credentials issued by one server are never reused against another. Authorize again to bring it back online.

## Troubleshooting

| What you see | What it means |
| --- | --- |
| **Authorization did not finish** | Click **Try again**. If you denied access, approve it on the next attempt. For another failure, contact support and include the error code shown in the message. |
| **Connected, but loading tools failed** | Authorization succeeded but the tool list did not load. Click **Reconnect**. |
| **No tools were found on this server. Try reconnecting.** | The server authorized but published no tools. Confirm the account you approved has access to the data you expect. |
| **Enter a valid URL, starting with https://** | The URL could not be read. Use the server's full `https://` address. Quickchat always connects over `https` so credentials are never sent unencrypted. |
| The action switched itself off | Its authorization expired or was revoked at the provider. Open the action and click **Reconnect**. |

:::note
MCP actions count toward your plan's Custom Actions quota, the same as API Actions.
:::