ChatGPT Data Access Requires Separate Plugin and Connector Setup
OpenAI’s workspace setup guidance separates enabling ChatGPT’s Data plugin from configuring the specific connections a team can use. In the walkthrough, Alistair Gillespie explains how admins can set installation access, publish a Snowflake app or create a custom MCP plugin, and review the risks of connecting an organization’s server. He also recommends connecting team messages, email and files to give ChatGPT more context for its answers.

Workspace access and connector setup are separate decisions
? alistair-gillespie distinguishes two admin choices that can otherwise blur together: enabling the Data plugin for the workspace, and configuring a particular connection that teammates can use. On the admin console’s Plugins page, an administrator can enable Data and choose whether teammates may install it themselves or whether it is installed for the workspace. The screen labels those options “Available for installation” and “Installed for workspace”; the latter can be uninstalled only by an admin.
Neither choice configures a data source. Enabling Data makes the plugin available under the selected workspace policy. A Snowflake connection, for example, still has to be set up as a separate, organization-specific app before teammates can install it and sign in. That distinction matters when planning rollout: workspace access determines who can get the plugin, while the connector setup determines which specific service it connects to.
A verified template does not mean the configured server has been reviewed
For warehouse connections such as Snowflake and Databricks, Gillespie points administrators to templated plugins. In the Snowflake example, the admin selects the template and enters organization-specific details, including a name, description, and server URL from the Snowflake admin. The screen’s example is named “Sales Analytics” and described as “Acme sales by customer and region.” It also presents authentication and advanced OAuth settings.
The template creates a workspace-specific draft app that can be published afterward. Gillespie says to review the sign-in settings and create the app; once published, the admin can make it “Available for installation,” so teammates can install it and sign in. In other words, enabling the Data plugin does not itself make this Snowflake connection ready: the organization must configure and publish the specific app, then set its availability.
The setup screen draws a security distinction between the template and the server entered by the organization. It says the Snowflake template is verified by OpenAI, but that OpenAI has not reviewed the specific MCP server. Its warning says an untrusted server may access or steal information shared through app use, or trick ChatGPT into using tools in unintended ways, including changing or deleting data. The screen asks the admin to acknowledge that risk before proceeding. Gillespie’s instruction is to review the sign-in settings; the screen makes clear that reviewing the template is not the same as reviewing the configured server.
A team-built server has its own creation path and risk review
For a connector or MCP server built by the team, the admin can use “Create MCP plugin” from the Plugins page rather than selecting the Snowflake template. Gillespie’s instructions are to give the plugin a name, paste the server address, choose how people sign in, review the settings, and create it. The on-screen example is “Acme Order Operations,” described as a way to track orders, shipments, and delivery exceptions.
The custom-plugin screen lists OAuth, access token or API key, and no authentication as sign-in choices; it also shows mixed authentication and advanced settings. It warns that custom MCP servers introduce risk. The Snowflake template and a custom MCP plugin are therefore different setup routes, but both require the admin to supply connection-specific settings and review them before creating the app.
Workplace context is part of the recommended setup
Gillespie recommends connecting more than warehouse data. He says the Data Agent is at its best when it can access a team’s messages, email, and files through plugins, and urges administrators to work with IT to set those connections up for the whole team.
ChatGPT's Data Agent is at its best when it has access to your team's messages, email, and files via plugins.
The accompanying guides cover workspace plugin access, the Snowflake app template, and custom MCP servers.