Source: Databricks Blog — 2026-08-04
Summary
Databricks announced Unity AI Gateway is now generally available, positioning it as a single control plane that routes and governs both AI model requests and MCP server traffic — enforcing rate limits, cost controls, and service policies (guardrails) while recording usage, access, and lineage across any model provider or coding agent. MCP servers, models, and functions are governed as Unity Catalog securables using the same privilege and policy system Databricks already applies to data tables. Databricks says thousands of customers — including Rivian, Asana, and Edmunds — already route AI traffic through the gateway, with over a quadrillion tokens passed through it in the past year.
Key Takeaways
- Unity AI Gateway routes every model request AND every MCP server request through one control plane — a notable scope expansion beyond a typical "LLM gateway" that only proxies model calls.
- It enforces rate limits, cost controls, and service policies (guardrails) at the routing layer, alongside end-to-end observability of usage, cost, access, and lineage.
- MCP servers are governed as Unity Catalog securables, meaning the same permission and policy model Databricks uses to govern data tables now extends to governing which agents/tools can reach which MCP servers.
- Headline adoption figure: Databricks reports over a quadrillion tokens have passed through Unity AI Gateway in the past year, across thousands of customers including named accounts Rivian, Asana, and Edmunds.
- Rollout is incremental — some adjacent capabilities, such as certain Unity AI Gateway service policies, remain in beta and require an account admin to enable them via the account console's Previews page even as the core gateway reaches GA.
Reel Script
Hook (~18s, ~40 words) Every AI agent in your company is now also an API client, a spending account, and a potential security hole — usually with zero central oversight. Databricks just took their data governance playbook and pointed it at that exact mess.
Core Concept (~70s, ~155 words) The core idea is simple to state and hard to actually build: put every AI-related request — not just calls to a model, but calls to MCP servers too — through one gateway that can see and control all of it. Quick refresher on why that matters: MCP servers are what give an agent live access to real systems, like your CRM or your internal database, so an ungoverned MCP connection is basically an ungoverned system access. Most "AI gateway" products only watch model traffic. Databricks is explicitly widening the net to cover MCP traffic too, then plugging the whole thing into Unity Catalog — the same permissions system Databricks already uses to control who can touch which data table. So a model, an MCP server, and a table all get governed the same way: same privilege model, same audit trail, one console.
Hands-On (~55s, ~125 words) The number Databricks is leading with is the concrete proof this isn't a paper launch: over a quadrillion tokens have already passed through Unity AI Gateway in the past year, across thousands of customers, with Rivian, Asana, and Edmunds named specifically. That's traffic flowing through rate limits, cost controls, and policy guardrails the gateway enforces at the routing layer, with lineage and access logged centrally instead of scattered across whichever tool each team happened to pick. Worth noting Databricks is upfront that this isn't fully finished — some of the newer service-policy features tied to the gateway are still in beta and need an admin to switch them on. GA is the core routing and governance layer, not every planned feature.
Takeaway (~20s, ~48 words) The real signal isn't the token count, it's the scope — governing MCP servers with the same rigor as data tables is the right instinct as agents get real system access. If your org has agents calling tools with nobody watching, this category is worth evaluating now, not after an incident.