Managing and Revoking Connections

How to review who has connected an AI assistant, revoke access, and understand the security model.

Reviewing authorized connections

Every connected AI assistant appears in the Authorized connections table on the API & MCP settings page, showing:

• Client - the application that connected (for example Claude, or your organisation's command-line client)

• Administrator - the administrator who approved the connection

• Status - whether the connection is pending, active, or revoked

• Last used - when the assistant last accessed your data

If several administrators connect the same app, each gets their own row - connections are always per person.

Revoking a connection

Click Revoke next to any active connection and confirm. The connection's access tokens stop working immediately; the next time that assistant tries to access your data it will be rejected and will need to be re-approved from scratch.

Revoking one administrator's connection does not affect anyone else's.

Disabling MCP access entirely

To switch off AI assistant access for your whole organisation, set Enable hosted MCP access to Disabled on the API & MCP page and save. All requests are rejected immediately and existing connections are revoked. You can re-enable access later, but every assistant will need to be reconnected and re-approved.

How security is enforced

• Administrator sign-in and consent. No assistant can connect without a Tahua administrator signing in through the browser and explicitly approving read-only access. There are no shared secrets or standing credentials - the connection configuration alone grants nothing.

• Read-only by design. The MCP server exposes no tools that create, modify, or delete data.

• Data policies apply. Every query is filtered through the connecting administrator's data policies, exactly as if they were browsing Tahua themselves.

• Every request is verified. Each request checks that the token is valid, the connection is still active, MCP access is still enabled for the organisation, and the administrator's account is still active. If any check fails, the request is rejected and the connection is revoked.

• Rate limiting. Requests are rate limited per connection to prevent abuse.

• Audit trail. Enabling or disabling access, provisioning a client, approving a connection, and revoking a connection are all recorded as security events.

FAQs

Does the AI assistant provider train on our data?

Tahua only makes your data available to the assistant you connect, under your organisation's agreement with that provider. Whether the provider uses conversation data for training is governed by your plan and settings with them - for example, Anthropic's commercial plans do not train on customer data by default. Check your AI provider's data policy before connecting.

Can applicants or assessors connect an assistant?

No. Only administrators can approve a connection, and connecting requires an administrator sign-in. Enabling and managing MCP access additionally requires the Manage Integrations permission.

Is this the same as API access?

They are related but separate. API Access provides OAuth credentials for your own systems to make programmatic REST requests. MCP access lets an AI assistant query your data conversationally, scoped to the individual administrator who connected it. Both are managed from the same API & MCP settings page.