For the engineer in the room

The whole product, front to back.

Airbrx speaks the protocols your applications already speak, transactional and analytical alike. One data path, any workload.

YOUR NETWORK AIRBRX NETWORK YOUR ACCOUNT Your LLM & agents connect over MCP MCP Your tools BI · dbt · JDBC · apps AIRBRX GATEWAY Gateway · MCP services on the path · SaaS-hosted deployed inside your network Your warehouse Snowflake · Databricks · Postgres policy Airbrx Admin App API · rules · config · reports Airbrx-hosted · off the data path deployed in your environment Segregated storage encrypted at rest Your storage in your own account cache · logs · rollups ON THE PATH — SEGREGATED & ENCRYPTED Airbrx doesn’t read your data — access governed by the DPA’s controls AIRBRX’S NETWORK: NOT ON THE DATA PATH queries and results never leave your network — only policy sync reaches Airbrx THE WHOLE STACK IN YOUR ENVIRONMENT admin app included — the data path and the control plane are both yours
SaaS: the gateway and the Airbrx Admin app run on Airbrx’s network. The gateway is on the path — and everything it keeps is segregated per tenant and encrypted at rest, with access governed by the DPA’s controls — Airbrx doesn’t read your data. Hybrid: the gateway, your data and your storage stay inside your network — Airbrx’s network is never on the data path. The Admin app stays Airbrx-hosted and syncs policy off the query path: the same control experience, with none of your query data. Custom: the whole stack — Admin app included — deploys inside your perimeter. The data path and the control plane are both yours. The green boundary is your network, the orange one is what Airbrx hosts — watch them trade territory as you move right through the tabs. The API is part of the Admin App, and your LLM and agents speak MCP straight to the gateway — both admin surfaces, both off the data path. Routing is per connection string in every model: only the tools you point at the gateway go through it at all.

The core

Six things, and nothing else on the path.

Rules engine

Serve it, or turn it away — by caller, by time of day, by identity, before it reaches the engine. Eight rules, worked through — in the exact schema the gateway runs.

Deterministic keys

Same inputs, same key. A hit returns byte-identical bytes, not a re-derivation.

Surgical invalidation

One rule’s results, never the whole cache. A pipeline call, a schedule, a rule edit or a write — each busts the subtree it should and nothing else.

Stale-while-revalidate

Answer now, refresh behind — where you have declared that trade acceptable.

Credential pass-through

The caller’s own token. Your row-level security and masking apply exactly as they do today.

Audit and rollups

One record per query, in a fixed shape — in your own bucket on a hybrid install, segregated and encrypted on SaaS.

Where it runs

Three deployment models.

SaaS, hosted by Airbrx

Nothing to run, and the lower-cost path. The gateway is on the path here — and everything it keeps is segregated per tenant and encrypted at rest, with access governed by the DPA’s controls — Airbrx doesn’t read your data.

Hybrid / cloud install

The gateway is deployed inside your network, so Airbrx’s network is never on the data path — queries and results never leave yours. Only policy sync reaches us, off the query path.

Custom install

The whole stack — Admin App included — deploys inside your perimeter. The data path and the control plane are both yours.

In every model, queries execute on your engine, under the caller’s own token — and the admin plane never sits on the query path.

Measured

One production run, with its limits named.

A real run against a real Databricks tenant, and an honest account of what it does and does not prove.

0
warehouse executions across the entire run
25/25
served from cache — every arrival, warm-up included
178ms
median, served — 187ms end to end

The run. Databricks gateway tenant, 29 July 2026. Open-loop, 60 queries a minute for 25 seconds, 2 distinct queries, 25 arrivals. 20 measured (5 warm-up excluded), 0 failed, 0 dropped, target rate achieved.

What it does not prove. The 100% hit rate is the workload’s design — two distinct keys — and is not a hit-rate result; do not read it as one. Latency here tracks driver roundtrips rather than cache status: 175ms median at one roundtrip (n=13), 627ms at four (n=7), and every one of those was a hit. With n=20 the tail is meaningless, so we quote the median and the roundtrip split and leave p95 alone. There was no direct-to-Databricks arm, so this is not a before-and-after. The cache was warm at the first arrival, so the run never exercised a miss-and-fill.

Don’t take our run for it. Get your own numbers.

The scan reads your own query history and tells you what the repetition is costing you — on your traffic, against your own bill.