--- description: Turn on the AI Analyst MCP server for your organization, choose which external AI apps may connect, and understand what those apps can reach. --- # MCP server The **MCP server** page in Settings controls whether members of your organization can use their AI Analysts from external AI apps such as Claude, ChatGPT and Cursor. You turn the server on, choose the apps you allow, and share one address with your members. For what the MCP server is for and what it can do, see [AI Analyst MCP Server](../mcp-server/README.md). !!! info "Availability" - **Who can do this:** members with the **Manage organization** capability. Every member can read the address and the allowed apps on [Connected apps](../agent/working-with-agents/connected-apps.md). - **Deployment:** Actian SaaS. Support for self-managed deployments is coming soon. If you don't hold **Manage organization**, **MCP server** does not appear in your Settings menu. ## Turn on the MCP server 1. Open **Settings** (from Studio or Explorer). 2. Under **Organization**, select **MCP server**. 3. Turn on **Enable MCP server**. This applies to the whole organization. The server is off by default, and every app is off by default. Turning the server on does not let anyone connect until you also allow at least one app. ## Allow apps The **Apps** section lists every app that can connect. Turn on the ones your organization allows. | App | Badge | | --- | --- | | Claude | **Draws charts** | | Claude Code | | | ChatGPT | **Draws charts** | | Codex | | | Cursor | **Draws charts** | | VS Code | **Draws charts** | | Zed | | | OpenCode | | | Goose | **Draws charts** | **Draws charts** marks apps that can draw charts and tables inside the chat. In the other apps, members get a text preview and a link that opens the chart or table in AI Analyst. Only the apps in this list can connect. An app you haven't turned on is refused at sign-in, and the member sees a message they can copy and send to you, asking you to turn it on. ## Share the address The **Address for your apps** section shows the **Server address** of your organization's MCP server. It has this form: ``` https:///mcp ``` Click **Copy link** and send it to your members, or point them to **Settings** → **Connected apps**, where they can copy it themselves and find setup guides for each allowed app. See [Connected apps (MCP)](../agent/working-with-agents/connected-apps.md). The address is not a secret. It does nothing until a member signs in with their own account and allows the app. ## Turn off the server or an app When you turn off the server or a single app, AI Analyst asks you to confirm (**Turn off the MCP server?** or **Turn off** *<app>***?**). | You turn off | What happens | | --- | --- | | The MCP server | Every connection in your organization ends. Apps that were signed in can no longer ask questions. | | One app | Every connection from that app ends. Other apps keep working. | !!! warning Turning the server or an app back on does not restore the connections that ended. Each member has to connect the app again. ## Access and data ### Who can use it A member can connect an app when all of these are true: * The MCP server is on and the app is allowed. * The member holds the **Chat** capability (or higher) in at least one Workspace. * The organization is active. An organization whose trial has ended, whose access is on hold or that is pending deletion can't be used through the MCP server. ### What a connected app can reach The app acts as the member who connected it. It never gets more access than that member has in Explorer. * **AI Analysts:** only deployed AI Analysts that the member can chat with in Explorer. A member with the **Edit** capability does not reach undeployed AI Analysts through the MCP server. * **Conversations:** only conversations the member owns that were started in Explorer or from an external app. Slack and Microsoft Teams threads are not included. * **Artifacts:** saved Artifacts the member owns or that are shared with them, and unsaved Artifacts from conversations the member can open. The app can't change anything. It can't save, share or archive Artifacts, and it can't edit AI Analysts or the semantic layer. It never queries your data sources directly; every question goes through an AI Analyst. If a member loses their last Workspace capability, their connections end on the next request. Getting access back later does not restore them; the member connects again. ### What is sent to the app Answers, chart data and table rows that AI Analyst returns are sent to the external app and to that app's AI provider. After that, the app's own terms and data handling apply. Allow only apps that your organization has approved for the data your AI Analysts can reach. ### Sign-in and connections * Members sign in to AI Analyst the way they normally do, including through Enterprise SSO, and then allow the app on a consent page. The app never sees their password. * A connection belongs to one member, one app and one organization. * A connection lasts 30 days. After that, the member signs in again from the app. ### Messages and limits A question asked from an external app counts against your organization's **Messages** allowance in the same way as a question asked in Explorer. When the allowance is used up, the app receives the message-limit notice as the answer. See [Billing & Usage](billing.md#when-you-reach-a-limit). ## Review usage * **Explorer:** a conversation started from an external app appears in the member's conversation history, labelled with the app, for example *via Claude*. * **Monitoring:** in Studio, filter conversations by the **MCP** channel to see the ones that came from external apps. See [Traces](../ai-analysts/monitoring/traces.md). * **Audit:** these actions are recorded in the [audit log](../governance/audit-logs.md). | Action | Recorded when | | --- | --- | | `org.mcp_server.set` | The MCP server is turned on or off | | `org.mcp_client.set` | An app is turned on or off | | `mcp.connection.open` | A member allows an app | | `mcp.connection.revoke` | A connection ends, with the reason | | `turn.start` | A question is asked. The entry names the MCP source and the app. | | `artifact.access` | An Artifact is read or shown through the app |