Databricks Introduces Unity Gateway CLI to Centralize Coding Agent Control

Databricks on September 24, 2026 introduced the Unity Gateway CLI, a command-line tool that connects coding agents such as Claude Code and Codex to the company’s Unity Gateway, where administrators manage approved models, tools, and spending policies. In an official blog post, Databricks said the tool, run through the ug command, is meant to pair developer freedom to adopt new agents and models with centralized governance across thousands of developers.

The Rationale Databricks Cites

Databricks framed the release against rapid model turnover. The company said GPT-6, Claude Opus 5.5, Gemini 3.8, and Grok 4.7 had all shipped in the preceding six months alongside open-weight models including Kimi K3, GLM-5, and DeepSeek V4.1, and characterized the current pace as a new frontier model roughly every five days.

Databricks said that pace puts the teams that own developer tooling in a squeeze: standardizing on one provider risks missing better or cheaper options elsewhere, while supporting many agents and providers leaves access control, budgets, security policy, and usage logging scattered across separate systems. The company said organizations need both the freedom to adopt the best option as it appears and a consistent way to govern it across thousands of developers.

Central Configuration and Agent Rollout

Unity Gateway now gives admins one console for configuring coding agents: default models, MCP servers, skills, Smart Routing, and spending policies, managed under Govern, then Agent Configuration. After admins publish a configuration, developers connect to any supported agent with a single command. The CLI authenticates the user, connects the chosen agent to Unity Gateway, and applies the published configuration before the agent’s regular interface opens. Developers can switch between enabled agents without configuring each one separately, and organizations can deploy ug through device management so teams start with the same approved setup, with settings admins can lock.

When new agents and models become available, admins pick the default setup they want developers to start with and publish it to the whole team as one configuration change. On the developer’s next ug run, the CLI syncs the new configuration and starts the recommended agent with the chosen model. Admins can revise that choice as quality, cost, and team needs change, and can define more detailed policies such as cohort-based rollouts of new models. Admins can also publish shared MCP tools and skills across supported agents, giving developers approved tools and proven workflows inside their configured agents without setting up each one separately.

Supported Agents, Installation, and License Terms

According to the public databricks/unity-gateway GitHub repository, the CLI runs coding agents through Databricks AI Gateway, with configuration support for Codex, Claude Code, Gemini CLI, OpenCode, GitHub Copilot CLI, and Pi, plus MCP-server registration for Cursor Agent, which connects through MCP only. The tool requires Python 3.12 or later and uv, with npm needed when an agent CLI requires automatic installation, and installs with a single uv tool install command pointed at the repository.

The first time a developer launches a model-backed agent, ug asks for a Databricks workspace, signs the user in, and writes the agent’s local configuration; later launches reuse the saved workspace and credentials. Existing ucode commands keep working, and the Python package retains the ucode name for compatibility. Any file ug overwrites is first backed up, and the ug revert command restores those backups. Databricks MCP servers are registered locally as stdio servers that run ug mcp-proxy, and the proxy refreshes Databricks OAuth tokens from the user’s CLI profile. Unity Catalog skills can be registered as MCP tools or downloaded into local agent skill directories, and the command reference includes ug status, ug doctor, ug usage, and ug upgrade.

The repository is published under the Databricks License, which allows viewing, using, copying, modifying, publishing, and distributing the materials only for use within or connecting to Databricks Services, with redistribution conditions and automatic termination upon breach. Databricks’ quickstart documentation presents ug as the recommended of two connection paths, the other being manual endpoint and authentication setup, and covers in-session model switching with /model plus ug mcp add and ug skills add for attaching tools and skills.

Cost Controls, Tracing, and Reported Usage

Admins can turn on Smart Routing centrally so that simpler tasks run on cheaper models and more difficult tasks on more capable ones; Smart Routing picks a model for the main session and can pick a different one for work the agent delegates to a subagent. Developers remain in their preferred agents without choosing a model for every task. Databricks said its published Smart Routing evaluation delivered 35% cost savings on its internal coding benchmark.

Budget-aware defaults in Unity Gateway let teams recommend lower-cost agents and models once spending reaches defined thresholds. The recommendations apply when new sessions start and do not interrupt work already in progress, and developers can check spending and remaining budget with ug usage.

Unity Gateway can also record traces from coding agents that cover local tool calls and skill invocations. Admins enable tracing centrally, and ug configures supported clients to export the traces to the lakehouse’s unified trace table, where teams can use Genie to surface recurring tool failures, oversized responses, and other token waste, then fix the problems and measure the results. Databricks said its own engineering organization used Unity Gateway tracing with Genie One to find and fix seven MCP tool bugs, which it said saved an estimated $1.2 million per year in wasted AI spend and lost productivity.

Concurrence chief technology officer John Xing said the CLI gives his company a single controlled connection for all of its coding agents’ model and tool traffic, with attribution down to the individual user and access restricted to approved models and MCP tools. He said Concurrence’s coding agents have processed more than 61 billion input tokens across roughly 360,000 requests since rollout, with usage and spending visible centrally.

Databricks said it already uses the CLI to push new models and cost optimizations across its own engineering organization while leaving developer workflows unchanged. The company pointed new users to the Unity Gateway CLI quickstart and to its documentation for configuring coding agents in Unity Gateway.