Skip to main content
The MCP Server is a per-project connector for your project’s own data. Once enabled, your users and their AI agents connect from a tool like Claude, Cursor, or ChatGPT and work with that project’s data — governed by each user’s project role, and limited to a data-only set of tools. It is separate from the Archie MCP: the Archie MCP manages the platform; this one is scoped to a single project.

​
Enable it

1

Open the panel

Go to Backend → App Services → MCP Server.
2

Turn it on

Toggle the MCP Server on. The connection details and configuration appear below the toggle.
Enabling the MCP Server does not open your data to everyone. Every request still authenticates as a specific user and runs with that user’s project roles. Enabling it only makes the connector available; access is still governed by your permissions.

​
Server identity

Give the server a name, icon, and description. These are optional — leave them blank to use the defaults. Edit the fields and click Save.

​
Authentication

How users connect depends on the authentication providers your project has enabled. The panel adapts the connection instructions to match:
If your project uses an identity provider (Archie Auth, Google, Microsoft, and so on), connecting opens a login window. The user signs in with their project account — there is no token to paste. After sign-in, the assistant acts with that user’s project role.
If no authentication provider is active on the project, users can’t connect. Enable at least one under Authentication Providers first.

​
Tools

The MCP Server exposes a data-only tool set. Assistants use these to understand your API and work with your data — never to administer the platform. Clients group them as read-only or write:
The exposed set is the intersection of this built-in floor and your own selection: you can turn a tool off to withhold it, but you can’t add tools beyond the data-only floor. Administrative and platform tools are never available here.
To narrow the set, toggle individual tools off in the panel. An empty selection uses the full data-only floor.

​
Permissions

The MCP Server never widens access. A connected user acts with their own project role:
  • Role-Based Access decides which operations and tables the user can touch.
  • Row-level filters apply exactly as they do on the GraphQL API.
  • Mutations run only when the user’s role allows the write and the environment permits mutations.
This is the same authorization the GraphQL API enforces — the MCP is just another authenticated caller.

​
Next step

Ready to connect a tool? See Connecting an assistant.

​
FAQ

No. Every request authenticates as a specific user and runs with that user’s project role and row-level filters. The connector is data-only and never exposes platform or admin operations.
No. The Archie MCP manages the platform across all your projects. The MCP Server is scoped to one project and works with that project’s data.
Both, within their permissions. Writes go through execute_graphql_operation and are allowed only when the user’s role permits the mutation and the environment allows mutations. Read-only and write tools are grouped separately so clients can require approval for writes.
Most assistants show a letter avatar for a connector added by URL — the per-server icon mainly matters when you publish the connector to that assistant’s directory. The name is what assistants display; the description is sent to the model as context rather than shown as a label.
Yes. Enablement, identity, and the tool selection are stored per project and environment.