# An agent's access is approvable only when identity, scope, log and revocation each have evidence

GUIDE, read 13 Nov 2026

**The short answer.** Sign-off on an AI agent's connection to analytics data rests on four documents: the identity it runs as, what each call may carry, what each call writes to a log, and who can revoke it. Of the 16 ranked tools, Atlan, Alation, DataHub and Ataccama document per-user identity, while Select Star and Secoda document shared tokens.

## The approver's first question is whose identity the agent runs as

An agent connection runs as one of two things: the person who asked, or a service account that stands for the agent. The choice decides what every later control can see. A policy keyed on the person works only if the person's identity reaches the data layer. A shared account makes every policy and every log line read the same name.

Vendor documents split on this. Atlan says its MCP server forwards the caller's identity so persona, role and domain policies apply to every tool call, and that a service-account token must carry exactly one persona. Alation says two users on the same MCP server can see different tool lists and results. DataHub says Cloud MCP uses OAuth2 so each user connects with a personal login, with tokens scoped to that user. Ataccama says each MCP request runs under the user's own identity, with Amazon Bedrock as the exception, which uses a shared service account. Informatica documents a gateway pattern with service-to-service OAuth 2.0, a service identity that is not stated per end user.

Ask the vendor for the sentence in the docs that names the identity, and for the exceptions. Ataccama's Bedrock line is the kind of exception that never reaches a sales deck and decides an audit finding. On the exposure map, the agent stage reads documented for Atlan, documented for Alation, documented for DataHub and partial for Ataccama ONE. See [the agent column of the map](/exposure-map/agent/) for all 16.

## What the agent can be told is set per tool, not per connection

Connecting to a server and being allowed to use its tools are separate grants in some products. Snowflake documents that access to a managed MCP server does not grant access to its tools, and that each tool needs its own grant. Snowflake also warns that exposing SYSTEM_EXECUTE_SQL on a server lets the client bypass the agent's semantic views and verified queries. An approver should read the tool list, not the server name.

DataHub says its MCP mutation tools, which add tags, terms and owners, stay off unless a setting enables them. Atlan scores 7 out of 10 on the site's agent scope criterion and Databricks Unity Catalog 7; the criterion is described on the [method page](/method/).

Alation adds a limit worth recording. Its docs say tool approvals are not enforced over MCP and each client must enforce its own tool restrictions. If the approval you are giving depends on a client prompt asking a human, the server will not back it up.

## A per-call log is the evidence an auditor will ask for

Atlan says every MCP tool call is logged with tool, client, model, tenant and duration, and that access denials are logged with reasons. Databricks says its audit log writes one event per Genie One MCP tool call. Immuta says its audit model records query, authentication, policy, project and tag events in one structure, and that logs expire after 90 days unless exported to S3 or ADLS Gen2.

Two checks follow. First, ask for a sample log line from a denied call, since a denial is the case that proves a control ran. Second, ask the retention figure and who owns the export. Immuta's 90-day default is stated and therefore easy to plan around; a vendor that names no figure leaves the question open.

## Revocation depends on what the credential is

Immuta says access for an agent is ephemeral: a role is created in the data platform and deleted after the action. Revocation there is the default state. At the other end, Select Star's docs say API tokens have no expiration and the MCP connection uses one, and Secoda's server code forwards a single API token as a bearer credential on each tool call. For those, revoking one agent means rotating a token that every client shares.

The reasoned verdict for the approval form: a credential that never expires and is shared by clients cannot be revoked per agent, so approve it only with a rotation date and a named owner. Ask each vendor to show the revoke step on a test agent and the log line that proves the next call failed.

## Approval itself should be a written rule with a log

Immuta's release notes for 4 Nov 2026 describe dynamic review flows that approve, deny or route a request, including requests from agents for pre-approved, non-sensitive datasets (see the [news item](/news/immuta-dynamic-review-flows-agent-access-requests/)).

What to request is the rule, its editor list and the decision log. If the product approves by tag, ask which tags it reads and what happens to data with no tag yet.

## A checklist table turns the questions into evidence requests

Use one row per question. A blank cell in the vendor's answer is a finding, written as none found on 13 Nov 2026.

| Question | Evidence to ask for | Documented in the ranked set |
| --- | --- | --- |
| Whose identity does each call carry? | The doc sentence naming user or service account, plus exceptions | Atlan, Alation, DataHub, Ataccama (per user); Informatica pattern (service) |
| What can a call carry? | Tool list, per-tool grants, a denied-call example | Snowflake (per-tool grants), DataHub (mutation tools off by default) |
| What is logged per call? | A sample record with user, tool, outcome and reason | Atlan (tool, client, model, duration), Databricks (one event per call) |
| How long is the log kept? | Retention figure and export path | Immuta (90 days, then export) |
| Who can revoke, and how fast? | A test revoke and the next-call failure in the log | Immuta (ephemeral role); Select Star and Secoda use long-lived tokens |
| How is a request approved? | The rule, its editors and the decision log | Immuta (dynamic review flows) |

## Sources

- Atlan: MCP security: https://docs.atlan.com/product/capabilities/atlan-ai/references/mcp-security (read 13 Nov 2026)
- Alation: using the MCP server: https://documentation.alation.com/en/latest/agentstudio/integrations/mcp/using-the-mcp-server.md (read 13 Nov 2026)
- DataHub: MCP server guide: https://docs.datahub.com/docs/features/feature-guides/mcp (read 13 Nov 2026)
- Ataccama: DQ&C MCP server: https://docs.ataccama.com/one/latest/generative-ai/mcp-dqc.html (read 13 Nov 2026)
- Informatica: MCPs on AWS AgentCore: https://www.informatica.com/blogs/powering-agentic-workflows-on-aws-with-informatica-mcps.html (read 13 Nov 2026)
- Snowflake: managed MCP server: https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp.md (read 13 Nov 2026)
- Databricks: audit log: https://docs.databricks.com/aws/en/admin/account-settings/audit-logs (read 13 Nov 2026)
- Immuta: agentic data access: https://documentation.immuta.com/saas/agentic-data-access/introduction.md (read 13 Nov 2026)
- Immuta: Detect audit model: https://documentation.immuta.com/saas/govern/detect-introduction/index.md (read 13 Nov 2026)
- Immuta: release notes: https://documentation.immuta.com/saas/releases/index.md (read 13 Nov 2026)
- Select Star: API authentication: https://web.archive.org/web/2026/https://docs.selectstar.com/select-star-api/authentication (read 13 Nov 2026)
- Secoda: MCP server code: https://raw.githubusercontent.com/secoda/secoda-mcp/main/server.py (read 13 Nov 2026)

Last reviewed 13 Nov 2026. Canonical: https://aidatagovernancetools.com/guides/approving-an-agents-access/
