Our methodology

Sourcing rules

Every capability we record links to a source with a publisher, a source type and the date we accessed it. Vendor documentation and official product pages are primary sources. Analyst summaries, press coverage and third-party blogs are secondary and never on their own sufficient to mark something verified.

Verification states

We use a set of distinct states rather than a yes/no checkbox, because “we could not find it” and “it does not exist” are different facts — and because a native capability should not look identical to one delivered by a third-party partner or locked to an enterprise plan.

Verified — official documentation

The vendor's own docs describe the capability explicitly.

Verified — official product page

Stated on the vendor's marketing site but not yet in documentation.

Available via integration

Documented, but delivered through a third-party partner or plug-in rather than natively by the gateway.

Available, enterprise plan

Documented and native, but gated behind an enterprise or paid tier rather than the base product.

Vendor claim — unverified

The vendor asserts it; we could not corroborate it in docs or a demo.

Not publicly documented

We searched official sources and found nothing. This is not the same as absent.

Verification pending

Research in progress. We publish the gap rather than guess.

Scoring criteria

Scores are currently marked pending verification across the site. We will publish numbers only once every input for a platform has a verified state — a score built on gaps is worse than no score. When published, each score will break down as follows.

Security capability

Injection defense, PII and secret handling, output inspection, agent controls, audit logging.

Cost optimization

Compression, caching, routing, budgets and attribution — scored separately, not merged.

Infrastructure

Model coverage, deployment options, latency posture, failover and integration surface.

Ease of use

Time to first protected request, documentation quality and operational burden.

Transparency

Public pricing, published limitations and documentation depth.

Handling vendor claims

Performance figures — token savings, detection rates, latency — are recorded verbatim and labelled as vendor-reported with their stated test conditions. We describe them as potential outcomes, never as guaranteed or independently benchmarked results, unless we have tested them ourselves and said so explicitly.

Update cadence and corrections

Each profile shows a last-reviewed date. We re-check tracked platforms on a rolling basis and after major product announcements. If you are a vendor and something here is wrong or out of date, tell us what changed and where it is documented, and we will correct the record and update the review date.

Next: browse the directory or read what an AI gate is.