Databricks Unity AI Gateway vs LiteLLM (2026): Managed Governance vs Open-Source Proxy

Three shapes on the same ten axes. Unity AI Gateway (orange) and LiteLLM (violet) both peak on the traffic spokes — LiteLLM edges Gateway & Policy Control 9 to 8 and they tie at 8 on Observability & FinOps — then diverge where the operating models differ: Unity AI Gateway holds Guardrails 8 to 6 with first-party inline policies, while LiteLLM takes Deployment Sovereignty 9 to 3, since the Databricks gateway cannot leave Databricks. Both cluster low on the inventory and compliance axes — the gateway category's signature. Kosmoy (blue) trades a little raw gateway breadth for reach across inventory, compliance and agent containment. Read it as area: the two gateways compete spoke by spoke; the suite covers the web.

Managed operation. The gateway is a service inside the platform — no proxy to deploy, no Postgres or Redis to run, no upgrade cadence to own. LiteLLM's enterprise tier is a license key applied to your own deployment: the customer owns the infrastructure, upgrades and on-call (enterprise docs).

Unity Catalog governance and audit. Per-principal permissions, policy decisions and usage land as queryable Unity Catalog records and system tables in the same catalog that governs the data (cost observability docs) — one governance model for tables, models, agents and MCP servers, where LiteLLM's admin-action audit logs live in the proxy.

Service-policy and guardrail depth. First-party PII redaction, safety, keyword and topic filters, jailbreak and hallucination blocking enforced inline before routing, plus Contextual Service Policies (Beta) that allow, deny or require approval for agent actions (guardrails docs). LiteLLM ships built-in and policy-template guardrails but orchestrates third-party engines rather than shipping detection models — an 8-to-6 gap on the radar.

MCP as a securable. External MCP servers register as Unity Catalog securables with fine-grained permissions, tool filtering and centralized audit, plus managed integrations for Google Drive, Jira, Confluence, Slack, GitHub and SharePoint (MCP governance blog) — MCP governed in the same catalog as the data it touches, where LiteLLM's registry, strong as it is among open-source gateways, is a separate proxy-level construct.

Finance-grade cost tables. Per-request usage with custom tags, hourly external-model spend by provider, target model and requesting principal, and enriched billing records for attribution across users, teams, applications, models and providers (cost observability docs) — in the billing system finance already reconciles, not a dashboard beside it.

It runs anywhere, including air-gapped. Docker, Kubernetes, Helm or Terraform on your own infrastructure, with explicit air-gap support and 'no data leaves your environment' as the enterprise posture (enterprise docs). Unity AI Gateway runs only as a Databricks-managed service — no self-hosted, on-prem or air-gapped option is documented as of August 16, 2026 — a 9-to-3 sovereignty gap.

MIT open source at zero license cost. The core proxy is free and MIT-licensed (~53.6k stars, ~9.8k forks); there is no DBU meter on logging and no consumption bill for governance features in the core (BerriAI/litellm). Unity AI Gateway is proprietary, and payload logging and usage tracking bill DBUs.

No platform commitment, broader provider matrix. 100+ providers behind one OpenAI-compatible API with no Unity Catalog workspace required — the gateway serves a Snowflake shop, a bare-Kubernetes shop and a three-cloud estate equally, where Unity AI Gateway presumes the Databricks platform.

Budget enforcement at every level. Budgets and spend attribution per organization, team, project, key and tag, soft-budget alerts, Prometheus and OpenTelemetry export and tool-call tracing (enterprise docs) — granularity mapped to your org chart and exported to your own observability stack, where Databricks' budgets and tables attach to platform principals and workspaces.

Community velocity. Weekly stable releases, a large contributor community and a staged Rust migration targeting sub-1ms overhead rolling out through December 2026 (announcement) — against a gateway that reached GA in August 2026 with service policies, agent services and Smart Routing still in Beta.

Both products govern the traffic that opts in: an app points its base URL at the gateway and inherits budgets, guardrails and logs. Neither documents what a regulated enterprise is increasingly asked to prove — an estate-wide inventory of AI running outside the gateway, evidence against the EU AI Act, ISO/IEC 42001 or NIST AI RMF, or kernel-enforced containment for agents that act. Unity AI Gateway's inventory covers assets registered in Unity Catalog; LiteLLM's covers assets routed through the proxy; both stop at their own perimeter as of August 16, 2026. When those are the ask, the category changes — from gateway to control plane, the frontier mapped on the AI capability map.

Kosmoy includes the gateway both products are — one OpenAI-compatible policy point with guardrails, RBAC, budgets and logging — but wraps it in the layers a gateway leaves out: a risk-tiered inventory of every model, MCP server and agent across the estate; EU AI Act, ISO 42001 (aligned) and NIST AI RMF evidence built from registry state plus gateway logs; and kernel-enforced Action Capsule containment with a kill switch. The direct head-to-heads live at Kosmoy vs Databricks Unity AI Gateway and Kosmoy vs LiteLLM.

And the pairings are real, not hypothetical: “LiteLLM + Kosmoy” shows up where a team keeps its open-source proxy for experimentation while Kosmoy holds the inventory, compliance evidence and containment for production; “Unity AI Gateway + Kosmoy” shows up where lakehouse workloads stay governed in Databricks while Kosmoy governs the AI estate beyond it. The honest framing is not that Kosmoy out-gateways either — it is that a gateway covers two or three spokes, and a suite covers the web.

For a team whose problem genuinely is model traffic, pick on the operating model: Unity AI Gateway if Databricks is the estate and a managed, consumption-billed service beats running a proxy; LiteLLM if the gateway must be open source, run on your own infrastructure — including air-gapped — and cost nothing in license. Both are OpenAI-compatible, so applications move with a base-URL change; the sticky parts are Unity Catalog registrations on one side and the operational investment in the proxy on the other. Plenty of Databricks shops run LiteLLM today for the workloads that never touch the lakehouse.

For an enterprise that has to prove control over all of its AI — not just route it — the choice is not between these two gateways but between a point tool and a suite. Kosmoy coexists with either pattern: LiteLLM or Unity AI Gateway keeps governing the traffic it already fronts, while Kosmoy holds the estate-wide inventory, the EU AI Act / ISO 42001 / NIST evidence and the containment layer for what reaches production.