# MCP Permissions

Source: https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/permissions/

Connecting an external AI client to the Qualytics MCP server needs no setup by an administrator and no MCP-specific permission. Your user role decides which tokens you can create to authorize a client. Once connected, the client runs every tool as the owner of that token, so your team permissions on each datastore decide what it can do.

## What Each Role Can Do

| Action | Member | Manager | Admin |
| --- | --- | --- | --- |
| Generate a Personal API Token and use it to authorize a client | Yes | Yes | Yes |
| Disconnect a client by revoking your own token | Yes | Yes | Yes |
| Create a service token, so a client can act as a service account | No | No | Yes |

Connecting a client does not require an AI provider in Qualytics, so the **Manager** and **Admin** steps that AgentQ needs do not apply here. See [Personal Token Permissions](https://userguide.qualytics.io/settings/access-tokens/personal-token/deep-dive/permissions/) for everything else a role allows with tokens.

## Tokens That Can Authorize a Client

| Token | Can authorize a client | The client acts as |
| --- | --- | --- |
| **Personal API Token** | Yes | You, with your role and team permissions. |
| **Service token** | Yes | The service account the token belongs to, with its role and teams. |
| **SCIM Administration Token** | No | Not applicable. The authorization page refuses it. |

When your deployment does not allow API tokens, no token can authorize a client, and the authorization page answers **Invalid API token**.

## What a Client Can Reach on Your Behalf

A client never widens what the token owner can already reach. A user whose access is restricted to a team sees only that team's datastores through the client, the same as in the web app. Three further limits apply on top of that:

| Limit | What it decides |
| --- | --- |
| **Scope** | A client authorized with read-only access can run the tools that read data, but not the tools that change it. |
| **Team permissions** | Each tool checks your team permission on the datastore it touches, with the same rules as the rest of the platform. |
| **Modules** | A tool that belongs to a module your deployment does not include answers with a message saying so, instead of acting. |

The AgentQ access level, which controls what AgentQ's built-in chat may share with its AI provider, does not apply to external clients. Each client brings its own model, so what it receives follows your role, your team permissions, and the client's scope.

## Team Permissions Behind the Tools

External clients call the same tools as AgentQ, so each tool needs the same team permission. Reading assets, checks, and anomalies needs the **Reporter** team permission. Creating a check as a Draft needs **Drafter**, creating or changing an active check needs **Author**, and running operations or changing computed assets needs **Editor**. The **Admin** role has access to every datastore and is not subject to team permission checks.

For the full list, kind of tool by kind of tool, see [Team Permissions Behind the Tools](https://userguide.qualytics.io/agentq/deep-dive/permissions/#team-permissions-behind-the-tools). If you lack the permission for a tool, the client receives a message saying the action was not permitted.

## Actions a Client Performs

When a client creates or changes something in Qualytics, the change is recorded against the owner of the token that authorized it, not against a shared identity. The item's Timeline also records the client as a co-author, under the name the client registered with, so you can tell which changes came through an external client.

Writes take effect as soon as a tool runs, and the tools have no confirmation step of their own. Keep your client's approval prompt turned on for tools that change data, and review what it proposes before you approve it. See [Best Practices](https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/best-practices/).

## See Also

- **Introduction**

  What MCP is, what the Qualytics MCP server offers, and why it matters for data quality.

  [Introduction](https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/introduction/)
- **How It Works**

  How a client, the Qualytics MCP server, and the model work together on a request, and what the model sees.

  [How It Works](https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/how-it-works/)
- **Connecting External AI Clients**

  What you need before connecting a client, and the setup for each company's clients.

  [Connecting External AI Clients](https://userguide.qualytics.io/using-the-platform/mcp/connecting-external-ai-clients/requirements/)
- **Examples**

  Real scenarios built on the prompts the Qualytics MCP server ships with, from explaining an anomaly to judging whether data can be trusted.

  [Examples](https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/examples/)
- **Best Practices**

  Keep client connections secure and get reliable results from them.

  [Best Practices](https://userguide.qualytics.io/using-the-platform/mcp/deep-dive/best-practices/)

*Last updated: November 7, 2026.*