Databricks has released a command-line tool that lets admins govern which coding agents, models, and tools thousands of developers can use, while developers keep launching their preferred agent with a single command. The Unity Gateway CLI, announced on the Databricks blog, sits between the fleet of coding agents an engineering org actually runs and the model providers behind them, applying centrally published settings at the moment a developer types a command like `ug claude` or `ug codex`.
The mechanics are straightforward: admins configure agent defaults, MCP servers, skills, Smart Routing, and spending policies in one place, inside Unity Gateway's Govern → Agent Configuration screen. Once that configuration is published, the CLI handles authentication, connects the chosen agent to Unity Gateway, and applies the published settings before opening the agent's normal interface. Developers don't set up each agent separately, and they don't need to pick a model per task — Unity Gateway does model-tool routing centrally, and organizations can push `ug` through device management so every developer starts from the same locked, approved baseline.
Smart Routing is the piece doing the cost work under the hood: it matches model cost to task complexity, sending simpler work to cheaper models and harder work to more capable ones, and it can make separate routing decisions for a main session versus a delegated subagent. Databricks' own published Smart Routing evaluation reports 35% cost savings on its internal coding benchmark. Budget-aware defaults extend the same logic to spend — once usage crosses defined thresholds, Unity Gateway recommends lower-cost agents and models as the new default for future launches, without interrupting agents already running. Developers can check their own spending against remaining budget with `ug usage`.
The operational numbers Databricks cites come from two places: a customer and its own engineering org. John Xing, Chief Technology Officer at Concurrence, says that since rolling out the Unity Gateway CLI, "our coding agents have generated more than 61 billion input tokens across roughly 360,000 requests, with centralized visibility into usage and spend." Separately, Databricks says it used Unity Gateway's tracing feature together with Genie to find and fix seven MCP tool bugs, an exercise it estimates saved $1.2 million per year in wasted AI spend and lost productivity. Tracing captures local tool calls and skill invocations from coding agents and exports them to what Databricks calls the lakehouse's unified trace table, where Genie is used to surface repeated tool failures and oversized responses.
What's notably absent from Databricks' post is latency data, per-token pricing for the gateway itself, or any breakdown of what the 35% Smart Routing savings figure assumes in terms of task mix — it is presented as a single benchmark result, not a range across workload types. The company also doesn't disclose how the seven MCP tool bugs were distributed across agents or how much of the $1.2 million figure is spend versus productivity, leaving that estimate as a blended number rather than an auditable one. Teams evaluating this will want their own tracing data before trusting a vendor-reported savings percentage against their own workload.
The harder trade is architectural rather than financial: centralizing model, tool, and policy control through a single gateway is exactly the kind of chokepoint that makes governance possible and also makes the gateway itself a new dependency and a new latency hop between every developer and every agent they run. Locking settings at the device-management level solves the "scattered access and budget controls" problem Databricks describes in its own framing of the squeeze between standardizing on one provider and supporting many — but it also means that whatever outage or misconfiguration happens at the Unity Gateway layer now affects every coding agent in the fleet simultaneously, not just one integration.
For teams running coding agents across more than one provider, the actionable move this week is to separate the policy layer from the agent layer now, before fleet size makes that separation expensive to retrofit — a single governed routing path is only a win if the routing decisions are inspected, not just trusted at 35%.