Controlling AI access

Choose one account for each connection

ChatGPT, Claude, and other MCP tools use Loam's configured shared resource. During OAuth, choose My account or an eligible workspace Agent. The choice stays fixed for that authorization and can be changed only by creating another authorization.

Access comes from the selected account

Loam checks the selected account's current direct grants, permission groups, denies, capabilities, and MCP restrictions on every operation. When an Agent is selected, the related human's personal or group data authority is never added. If the Agent loses access, or the human loses permission to use the Agent, the next request fails or narrows.

Review connected tools

Open Settings > Connect AI (sign in to open this page) to review your own authorizations, selected accounts, related human, live access, and restrictions. Revoking one authorization does not revoke the shared MCP resource, an Agent, or another human's independently authorized connection.

Create scoped API access

For scripts and API clients, open Settings > API keys (sign in to open this page). An API key authenticates one human or Agent and may optionally narrow that account's access. Loam binds any narrowing rules to stable node IDs, not editable dataset or table names.

What each role allows

• Viewer reads data and metadata.
• Editor mutates rows and creates, changes, or drops ordinary tables and views. It cannot manage grants, publication, connections, or dataset structure.
• A human Admin on a node can manage dataset structure and grants for that node's subtree. API keys and other machine credentials cannot delegate grants, even when their data role is Admin.

How query access is checked

For each authenticated operation, the Engine evaluates the selected principal's current allow and deny rules and then intersects any authenticator ceiling. A deny is an absolute veto: for example, denying one table keeps it unavailable even when an ancestor dataset is allowed. Permissions inherit through every level of the dataset tree, and table listing, descriptions, queries, and changes all use that same ancestry and effective decision.

Best practices

• Start narrow. Select the least-privileged human or Agent suitable for the client.
• Use separate credentials for separate uses. This lets you update or revoke one without interrupting unrelated tools.
• Review periodically. Revoke connected tools and API keys that no longer need access.