# Whether an agent sees what the asking person may see depends on whether the server runs calls as that person or as one shared credential

GUIDE, read 10 Nov 2026

**The short answer.** Atlan, Cube, Hex, Omni and Tableau document calls that run as the signed-in user, while Fivetran uses one API key and secret, the Secoda wrapper one API_TOKEN, and Select Star one API token. Cube applies row-level security, Euno and OpenMetadata use personas, and Looker passes the caller's OAuth token only when use_client_oauth is switched on.

## OAuth on behalf of the user ties each call to a person

When a server completes OAuth for the person at the keyboard, the vendor can apply that person's permissions to every tool call. Per the docs for [the Atlan MCP server](/servers/atlan-mcp/), OAuth makes each tool call run as the signed-in user, so tools run with their own permissions and actions are attributed to them. Per the docs for [the Cube MCP server](/servers/cube-mcp/), every tool runs as the authenticated user, and queries respect Cube permissions including row-level security.

Per Omni's docs, MCP tools run with the permissions of a specific Omni user, so results include only what that user can access. Per Sigma's help page, the server inherits the user's Sigma account permissions, and access needs the 'Use Sigma MCP with OAuth' account-type permission plus at least Can view on documents or Can use on connections. Per Tableau's README, the hosted server uses OAuth 2.1 so each user signs in as themselves and per-user Tableau permissions are enforced.

Google documents OAuth 2.0 with IAM for the BigQuery remote server, so access follows the authenticated Google Cloud identity. Per Databricks, the dbsql MCP runs SQL using the caller's Unity Catalog and warehouse permissions, so results are scoped to the asking user.

## Service accounts and shared keys scope the agent, not the person

A shared credential gives every asker the same view. Per the README of [the Fivetran MCP Server](/servers/fivetran-mcp/), the server uses one Fivetran API key and secret, so agent actions run with that credential's scope instead of the asking user's identity, and OAuth resource-server support in streamable-http mode is not yet functional. Secoda's docs say the open-source wrapper authenticates with a single API_TOKEN, one shared token per server process, while the Secoda-hosted endpoint verifies the user's workspace and AI permissions.

Select Star's page describes an API token sent as a Bearer header. No per-user OAuth is described, and the page does not say results are limited to the asking user's permissions (none found on 10 Nov 2026). Per Collibra's README, stdio mode uses one server-wide credential, while HTTP mode advises that MCP clients supply their own Basic Auth headers per request.

Alation's guide describes the service-account route: the agent gets the data and features allowed by the role assigned to the service account. Per OpenMetadata's docs, unattended agents should use a Bot service account with a role-based access policy, kept separate from personal logins and auditable. Per Airbyte's tools page, direct reads use the credentials stored in the connector rather than the asking user, so source-side access is not per-user even though the tool calls are.

## Some servers offer both and the default decides the outcome

Looker's source page says the Toolbox can pass the caller's OAuth token instead of one service credential via use_client_oauth, which defaults to false. With the default, identity is the API user configured with client_id and client_secret. The same Toolbox documents useClientOAuth for BigQuery, which forwards the client's OAuth token so calls run as the asking user, and impersonateServiceAccount for service account impersonation.

Omni lets an Organization API key run tools as another user through the X-MCP-User-ID header. Cube's API-key tokens can act as the key's owner, as another user, or be limited to specific tools, and last up to one hour with no refresh token. Tableau's self-hosted server offers PAT (the default), direct-trust, OAuth and passthrough authentication, per its configuration docs.

DataHub splits by transport. Per its README, the Docker HTTP deployment has each MCP client supply its own DataHub token, so permissions and audit identity are preserved per user, and the HTTP entry point refuses to start if a shared server-wide token is configured. Over stdio the README uses DATAHUB_GMS_URL and DATAHUB_GMS_TOKEN.

## Personas narrow what an agent is shown

A persona is a named profile of what a role should see. Per Euno's docs, the server uses the active persona, and if none can be identified the user is asked to choose; the docs say the active persona scopes data and actions, and list_personas, current_persona and switch_persona are among the published tools. Per Atlan's security page, guests and members are scoped to their persona, and MCP narrows but never widens what a user's login permits.

OpenMetadata's reference lists get_persona_context and find_context, which serve shared AI context documents curated per persona, and get_user_context returns the caller's identity, roles, teams, domains and active persona, with no parameter to look up another user. These tools shape context. The access boundary in each case is the platform's own policy, as the vendors describe it.

## Row-level security and role gates appear in the semantic and BI servers

Cube's docs name row-level security for MCP queries. Microsoft's overview says Fabric IQ, its consumption server, answers questions over semantic models and respects permissions, row-level security and object-level security. Microsoft's authoring server goes the other way: its tools manage security roles with row-level security inside the model, per the README.

Role gates are more common. Per Monte Carlo's docs, an Editor role or above is required to use the server. Hex requires the Explorer role for knowledge tools and Editor or above for project editing tools. Cube says a Viewer role or higher is needed and that data model editing tools show only for roles allowed to edit the model, Admin and Developer by default. Alation's remote server lists tools dynamically and filters out tools the current user cannot use. Snowflake's docs say access to the MCP server does not give access to tools, since privileges are granted per tool.

## The documented identity of each server, side by side

The table records the identity each publisher says a call runs as. The sources for each row are listed at the end of the guide. Where a server offers more than one route, both are shown.

A row that reads none found means the page read on 10 Nov 2026 did not state how results are scoped. Per-user scoping also depends on the client and on how the customer configures roles, so the vendor's statement is the starting point for a test and not the end of it.

| Server | Identity the call runs as | Per-user scoping documented |
| --- | --- | --- |
| Atlan MCP (hosted) | Signed-in user via OAuth, or API key as a shared identity | Yes, with OAuth |
| BigQuery remote MCP server | Authenticated Google Cloud identity via OAuth 2.0 and IAM | Yes |
| Collibra MCP Server | Server-wide credential over stdio; client Basic Auth headers over HTTP | Per request over HTTP |
| Cube MCP server | Authenticated user, or API-key owner or chosen user | Yes, with row-level security |
| Fivetran MCP Server | One API key and secret | Not documented |
| Looker tools in MCP Toolbox | API user, or caller's OAuth token if use_client_oauth is true | Opt-in |
| Select Star MCP Server | API token as Bearer header | none found on 10 Nov 2026 |
| Tableau MCP | Hosted: each user's Tableau Cloud identity; self-hosted: PAT by default | Hosted only |

## Sources

- Atlan remote MCP overview: https://docs.atlan.com/product/capabilities/atlan-ai/how-tos/remote-mcp-overview (read 10 Nov 2026)
- Atlan MCP security page: https://docs.atlan.com/product/capabilities/atlan-ai/references/mcp-security (read 10 Nov 2026)
- Cube MCP server docs: https://docs.cube.dev/docs/integrations/mcp-server.md (read 10 Nov 2026)
- Omni MCP tools: https://docs.omni.co/ai/mcp/tools.md (read 10 Nov 2026)
- Sigma MCP server help page: https://help.sigmacomputing.com/docs/use-sigma-mcp-server (read 10 Nov 2026)
- Tableau MCP repository: https://github.com/tableau/tableau-mcp (read 10 Nov 2026)
- Tableau MCP environment variables: https://tableau.github.io/tableau-mcp/docs/configuration/mcp-config/env-vars (read 10 Nov 2026)
- BigQuery remote MCP guide: https://docs.cloud.google.com/bigquery/docs/use-bigquery-mcp (read 10 Nov 2026)
- Databricks managed MCP page: https://docs.databricks.com/aws/en/generative-ai/mcp/managed-mcp (read 10 Nov 2026)
- Fivetran MCP README: https://raw.githubusercontent.com/fivetran/fivetran-mcp/main/README.md (read 10 Nov 2026)
- Secoda MCP docs: https://docs.secoda.co/features/ai-assistant/secoda-mcp-server.md (read 10 Nov 2026)
- Select Star MCP docs: https://docs.selectstar.com/features/mcp-server (read 10 Nov 2026)
- Collibra MCP README: https://github.com/collibra/chip/blob/main/README.md (read 10 Nov 2026)
- Alation authentication guide: https://github.com/Alation/alation-ai-agent-sdk/blob/main/guides/authentication.md (read 10 Nov 2026)
- Alation agent SDK README: https://github.com/Alation/alation-ai-agent-sdk/blob/main/README.md (read 10 Nov 2026)
- OpenMetadata MCP guide: https://docs.open-metadata.org/v2.0.x/how-to-guides/mcp.md (read 10 Nov 2026)
- OpenMetadata MCP reference: https://docs.open-metadata.org/v2.0.x/how-to-guides/mcp/reference.md (read 10 Nov 2026)
- Airbyte MCP tools: https://docs.airbyte.com/platform/airbyte-mcp/tools.md (read 10 Nov 2026)
- MCP Toolbox Looker source: https://mcp-toolbox.dev/integrations/looker/source/ (read 10 Nov 2026)
- MCP Toolbox BigQuery source: https://raw.githubusercontent.com/googleapis/mcp-toolbox/main/docs/en/integrations/bigquery/source.md (read 10 Nov 2026)
- DataHub MCP README: https://raw.githubusercontent.com/acryldata/mcp-server-datahub/main/README.md (read 10 Nov 2026)
- Euno MCP setup docs: https://docs.euno.ai/ai/mcp-setup.md (read 10 Nov 2026)
- Microsoft Learn MCP servers overview: https://learn.microsoft.com/en-us/power-bi/developer/mcp/mcp-servers-overview (read 10 Nov 2026)
- Power BI Modeling MCP README: https://raw.githubusercontent.com/microsoft/powerbi-modeling-mcp/main/README.md (read 10 Nov 2026)
- Monte Carlo MCP server docs: https://docs.getmontecarlo.com/docs/mcp-server (read 10 Nov 2026)
- Hex MCP server docs: https://learn.hex.tech/docs/api-integrations/mcp-server (read 10 Nov 2026)
- Snowflake-managed MCP docs: https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://datamcpservers.com/guides/scoping-context-to-the-person-asking/
