Skip to main content

Set up model capacity

Make models available to your team's coding agents through Unity Gateway in two ways: use Databricks-provided models, or connect your organization's own model provider accounts. You can offer both, then grant your team access to the models they need.

Use Databricks models

Your workspace's Databricks model APIs are available in Unity Gateway, referenced by their Unity Catalog names, such as system.ai.claude-sonnet-4-6. These are the foundation models your workspace serves.

To make a model available to your team, grant access to its model API (see Grant access to models, MCP servers, and skills), then add the model as a source for an agent in your coding agent configuration.

For the available models and setup details, see Govern model APIs.

Bring your own model providers

Connect coding agents to your organization's OpenAI, Anthropic, or Amazon Bedrock account through Unity Gateway. Register the external model provider once, then let developers use it with their Databricks identity. Unity Gateway supplies the provider credentials, so developers do not need their own copy of the API key.

Register a provider

In Unity Gateway, open Providers and create a provider. Choose its type, catalog, and schema, then enter its connection details, credentials, and the models to make available. For Amazon Bedrock, you can use a Databricks service credential instead of an access key pair.

The provider has a Unity Catalog name such as main.default.anthropic_prod. See Create and manage external providers for setup steps and authentication requirements.

Use the provider

To distribute the provider to your team, select it as the model source for an agent in your coding agent configuration. Developers apply the configuration with ug configure and launch the agent as usual.

To select a provider when launching an agent, pass its Unity Catalog name to --provider:

Agent

Supported provider

Launch command

Claude Code

Anthropic or Amazon Bedrock

ug claude --provider main.default.anthropic_prod

Codex

OpenAI

ug codex --provider main.default.openai_prod

Gemini CLI

Google Gemini Enterprise

ug gemini --provider main.default.gemini_prod

Agent

Supported provider

Launch command

Claude Code

Anthropic or Amazon Bedrock

ug claude --provider main.default.anthropic_prod

Codex

OpenAI

ug codex --provider main.default.openai_prod

Gemini CLI

Google Gemini Enterprise

ug gemini --provider main.default.gemini_prod

Replace the example names with your registered external model providers. The command opens the agent with requests routed through that provider. Select a model the provider makes available; for Amazon Bedrock, use its provider-specific model IDs.

These are the provider combinations supported by the ug launch commands with --provider. The Unity Gateway full provider list includes additional external model providers.

Grant access to models, MCP servers, and skills

Adding a resource to a coding agent configuration does not grant access to it. Grant developers the required Unity Catalog privileges before distributing the configuration.

Resource

Required privilege

Details

Models

EXECUTE on the model API.

Govern model access

External model providers

EXECUTE on the provider.

Govern external provider access

MCP servers

EXECUTE on the registered MCP server. Users also need any permissions required by its underlying tools and data.

Govern MCP access

Skills

READ VOLUME to read and download the skill.

Govern skills

Resource

Required privilege

Details

Models

EXECUTE on the model API.

Govern model access

External model providers

EXECUTE on the provider.

Govern external provider access

MCP servers

EXECUTE on the registered MCP server. Users also need any permissions required by its underlying tools and data.

Govern MCP access

Skills

READ VOLUME to read and download the skill.

Govern skills

Users also need USE CATALOG and USE SCHEMA on each resource's parent catalog and schema. Manage these grants for your team's group in Catalog Explorer.

Use Unity Catalog permissions to restrict access. Removing a resource from the configuration changes the shared setup; it does not revoke an existing grant.

Next steps