AI Gateway
in/guard/out vs Databricks Unity Gateway
Databricks Unity Gateway – formerly AI Gateway, generally available since August 2026 – is an enterprise AI gateway built on Unity Catalog. It routes traffic across proprietary and open models, governs coding agents, MCP servers, tools and skills, and brings it all under one control plane: budgets and hard spend caps by user, team or workspace, rate limits, identity-aware service policies that can allow, deny or require approval for an action, guardrails for PII, prompt injection, jailbreaks and unsafe content, and end-to-end tracing into governed tables. For a company already on Databricks, it is the natural choice for the estate.
in/guard/out works at a finer grain: the single agent run. It checks that the numbers in an answer match the source, that a payment’s IBAN came from something the run read, that a screened value never leaves through a tool argument, and that one task stays inside its own budget – and when something is wrong, it repairs the output instead of only blocking it.
Side by side
| Capability | Databricks Unity Gateway | in/guard/out |
|---|---|---|
| Wrong numbers in an answer | Custom guardrails judged by an LLM | Deterministic grounding: arithmetic recomputed and altered figures corrected from the source, no model call |
| Private data | Detected and masked to labels like [NAME] and [EMAIL], or blocked | Stable placeholders, restored in the response |
| When a check fails | Block, sanitize, or log | Repair where safe: the number, the JSON, the PII |
| Agent spend | Budgets and hard caps per user, team, use case, or workspace | Those still apply, plus per-run cost, step, and tool-call caps reserved before each call |
| A payment’s payee and amount | Policies on the tool, the user, and request content | Must appear in something the run read – an injected payee that matches no source is refused |
| Screened data leaving through tools | Not described publicly | Taint tracking follows screened values into tool arguments |
| Output format (JSON schema) | Not described publicly | Validate and repair |
| Runaway loops | Rate limits per user, team, or service | Identical tool calls fingerprinted across the run and stopped past a threshold |
| Workflow intent | Identity- and attribute-based service policies | Tool-ordering rules compiled from an n8n export and enforced across the run |
| Where it runs | On Databricks, billed per second of usage | In front of any provider, OpenAI- and Anthropic-SDK-compatible |
Based on Databricks Unity Gateway’s publicly available information. We’ve done our best to keep it accurate, but details may have changed since we last looked. If you’ve noticed an error, please email us.
Governing the estate and checking the run
Unity Gateway answers the platform team’s questions: which models and tools may each person, team and agent use, how much may they spend, and what did they do. It answers them well, and for traffic that already lives on Databricks it ties AI usage to the same catalog that governs the data.
in/guard/out answers the questions that come up inside one task. Is the figure in this report the one in the source? Did the account number the agent is about to pay appear in anything it read, or did a document plant it three steps ago? Has this run, not this user’s month, already spent what it should? Those checks need the run’s own history in hand, and most of them are deterministic – no judge model, no token cost. When one fails, in/guard/out repairs what it safely can and tells the agent how to fix the rest.
The two compose. Unity Gateway keeps governing the estate – access, spend, policies, traces – while the agents whose actions touch numbers, money or customer data run through in/guard/out and get checked step by step.
Frequently asked questions
Is in/guard/out a Unity Gateway alternative?
Not for the platform job. Unity Gateway governs model and tool access across an organisation, with spend caps, rate limits, service policies and tracing built on Unity Catalog. in/guard/out checks individual agent runs: grounding, PII restore, payee checks, per-run budgets and repair. If you’re on Databricks, keep Unity Gateway and add in/guard/out where agents act on numbers, money or customer data.
Unity Gateway has guardrails – what does in/guard/out add?
Unity Gateway’s guardrails screen PII, injection, jailbreaks and unsafe content, and run custom checks judged by an LLM. in/guard/out adds checks that need no model: figures matched to the source, arithmetic recomputed, a payee matched to what the run read. And it repairs what fails instead of only blocking it.
Both have budgets – do they conflict?
No, they nest. Unity Gateway caps spend per user, team, use case or workspace – the outer wall. in/guard/out caps one agent run: cost, steps and tool calls, guard-model calls included, reserved before each request so concurrent calls can’t overshoot.
Do I need Databricks to use in/guard/out?
No. in/guard/out fronts OpenAI, Anthropic, OpenRouter and local OpenAI-compatible models. You point your app’s SDK at it with a base-URL change.
Govern the estate. Check the run.
Unity Gateway governs who uses which model and tool. in/guard/out checks what each agent run actually does - FIX mode observes first, so you can measure it before enforcing. We are running a limited demo - sign up and we will get you in as soon as we can.