---
title: How do you control which Databricks users and agents can query model serving endpoints?
url: https://opencontroller.lyzr.com/blog/databricks-serving-endpoint-permissions-for-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do you control which Databricks users and agents can query model serving endpoints?

**Grant CAN QUERY on approved serving endpoints to the users, groups and service principals that need them, and keep CAN MANAGE and endpoint creation with admins. Then add AI Gateway rate limits per user, group or service principal and turn on usage tracking, so every agent identity has both an access boundary and a consumption boundary.**

In Databricks, foundation model and agent endpoints are workspace objects with their own access control lists. Getting those ACLs right matters more than any prompt-level control, because a user with CAN MANAGE can change the configuration that enforces everything else.

## What do the endpoint permission levels allow?

| Ability | NO PERMISSIONS | CAN VIEW | CAN QUERY | CAN MANAGE |
|---|---|---|---|---|
| Get and list endpoints | | ✓ | ✓ | ✓ |
| Query the endpoint | | | ✓ | ✓ |
| Update configuration | | | | ✓ |
| Delete the endpoint | | | | ✓ |
| Modify permissions | | | | ✓ |

The endpoint creator receives CAN MANAGE automatically. That is the main governance gap: anyone who can create endpoints controls their own guardrails and rate limits. Databricks' own guidance is to restrict endpoint creation and CAN MANAGE to admins and give other users only query permissions on approved endpoints, so guardrails and throughput limits cannot be bypassed.

## How should agents get access?

Agents should call endpoints as service principals, not as the developer who built them.

1. **Create a service principal per agent** (or per agent and environment) so access and usage are attributable.
2. **Grant CAN QUERY on specific endpoints** to that service principal. Avoid granting it on every endpoint "for convenience".
3. **Use groups for people,** for example a group of analysts allowed to query one approved chat endpoint.
4. **Keep CAN MANAGE with a platform team** that owns endpoint configuration.
5. **Review ACLs on a schedule** and remove service principals for retired agents.

Where an agent must act with a user's own permissions instead, see [Databricks on-behalf-of-user auth for agents](/blog/databricks-on-behalf-of-user-auth-for-agents).

## How do AI Gateway rate limits add a consumption boundary?

ACLs decide who can call. AI Gateway rate limits decide how much. On serving endpoints you can set limits in queries per minute (QPM) or tokens per minute (TPM):

- **Endpoint limit:** a global maximum; when exceeded, all requests are blocked.
- **Per-user default:** applies to every caller.
- **Custom limits** for individual users, service principals or user groups.

Constraints from Databricks' docs:

- Up to 20 rate limits per endpoint, including up to 5 group-specific limits.
- When both QPM and TPM apply, the more restrictive wins.
- TPM limits cannot be applied to endpoints serving custom models or agents, so use QPM there.

Databricks now documents a newer gateway experience as its recommended control plane alongside these endpoint-level features, so check which applies to your workspace. Our guide to [Databricks AI Gateway for agents](/blog/databricks-ai-gateway-for-agents) covers guardrails, traffic splitting and fallbacks.

## How do you audit who called what?

- **Usage tracking** writes token counts per request to `system.serving.endpoint_usage`, with endpoint metadata in `system.serving.served_entities`. It requires Unity Catalog, and only account admins can query these tables by default.
- **Inference tables** log requests and responses to Unity Catalog Delta tables. Payloads over 1 MiB are not logged. See [Databricks inference tables for agent audit](/blog/databricks-inference-tables-for-agent-audit).

Join usage to your service principal list to produce per-agent consumption reports.

## A setup checklist

1. Restrict endpoint creation to admins.
2. One service principal per agent; CAN QUERY on named endpoints only.
3. No CAN MANAGE outside the platform team.
4. Endpoint-level QPM ceiling plus per-service-principal limits.
5. Usage tracking and inference tables on for every agent-facing endpoint.
6. Quarterly ACL review tied to your agent inventory.

For the wider Databricks control set, read [governing Databricks agents](/blog/govern-databricks-agents), and for the cross-platform view see the [platform guides pillar](/blog/how-to-govern-ai-agents-across-platforms).

## How Opencontroller does this

Opencontroller connects to Databricks as a third-party agent platform and brings its agents into one governed registry with agents on AWS, Azure, Google Cloud, Salesforce, SAP and HubSpot. Posture detects risky permissions and configurations, agent workload identity federation and delegated on-behalf-of identity cover how agents authenticate, and per-agent Block and Quarantine act through the connected platform. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Who gets CAN MANAGE on a new serving endpoint?**
The endpoint creator receives it automatically. Restrict endpoint creation to admins if you do not want builders controlling their own guardrails and limits.

**Can I rate-limit a specific agent on a shared endpoint?**
Yes. AI Gateway supports custom QPM or TPM limits for individual service principals, users or groups. TPM is not available on endpoints serving custom models or agents.

**Who can read serving usage system tables?**
By default only account admins can query `system.serving.endpoint_usage` and `system.serving.served_entities`.

## Key facts

- Databricks serving endpoint ACLs: NO PERMISSIONS, CAN VIEW (get/list), CAN QUERY (get/list/query), CAN MANAGE (also update config, delete, modify permissions). The endpoint creator gets CAN MANAGE automatically.
- Databricks recommends restricting endpoint creation and CAN MANAGE to admins and giving others only query permissions to avoid bypassing guardrails or throughput limits.
- AI Gateway rate limits: QPM or TPM at endpoint level, per-user default, custom limits for users, service principals or groups; up to 20 limits, 5 group-specific; more restrictive of QPM/TPM wins; TPM not available for custom models or agents.
- Usage tracking writes to system.serving.endpoint_usage (account admins only); inference tables do not log payloads over 1 MiB.
- Opencontroller by Lyzr connects to Databricks and governs its agents in a cross-platform registry; pricing starts from $250K per deployment per year.

## Related

- [What does Databricks AI Gateway govern for agents?](https://opencontroller.lyzr.com/blog/databricks-ai-gateway-for-agents.md)
- [How do you govern AI agents built on Databricks?](https://opencontroller.lyzr.com/blog/govern-databricks-agents.md)
- [What do Databricks AI Gateway inference tables log for AI agent audit?](https://opencontroller.lyzr.com/blog/databricks-inference-tables-for-agent-audit.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
