Each team picks a model
SDK drift, key sprawl, and no shared failover when a provider degrades.
For developers, CTOs, CIOs, CISOs, and CFOs
Assert is the gateway between your applications and every model. It inspects the prompt, sends it only to an allowed model, and records who called, what rule applied, and what it cost. Cloud providers are optional. The same gate runs on your network, including airgapped.
Green continues. Amber is redacted, then routed. Red is denied and written to audit only.
Unhealthy providers trip circuit breakers. Failover stays inside the same residency and allow list.
Direct provider calls put keys in apps, secrets on the wire, and spend on an invoice with no owner.
SDK drift, key sprawl, and no shared failover when a provider degrades.
PII, credentials, and confidential content are inspected after the call, if at all.
Finance can see the provider bill. It cannot see which workload created it.
Applications change a base URL. Assert holds the provider keys and the retry policy.
Allow, redact, tokenize, or deny. A blocked credential never becomes a provider request.
Spend is cut by org, team, app, user, model, and provider, and can stop the next call.
Same gate. Different outcome. Developer is first because nothing else lands without the client change.
If the prompt can leave, or the bill cannot be named, this is the gate.
Account numbers and credentials are tokenized or blocked. Residency keeps the call in-region. Spend lands on the desk that ran it.
ePHI paths redact before send. Airgapped tenants can be limited to models you host. Access stays on the audit row.
Signed policy bundles, no license callback, and a candidate list that contains only local models.
One OpenAI-compatible base URL across teams. Failover when a provider degrades. Quotas before the surprise invoice.
Low-risk traffic routes for cost. A secret in the transcript denies the call instead of forwarding it.
Parent policy inherited by team and app. Units can narrow a rule. They cannot open a side door.
Composite scenes, not named customers and not measured results. Use them to recognize the room, not as a reference.
“We stopped arguing from the invoice. The deny row names the app.”
Composite · CFO and platform lead, multi-team software company“The credential case never produced a provider request. That was the control we could show.”
Composite · CISO review, regulated extraction workload“Apps changed a base URL. We did not rewrite the SDK layer.”
Composite · developer rollout“The enclave has no route to a commercial model unless policy adds one.”
Composite · airgapped deploymentAssert does not publish customer logos or outcome metrics on this page. A briefing can use your own volumes in the ROI calculator.
Four steps. The application keeps an OpenAI-compatible client.
Providers are adapters. Identity, detectors, and sinks are replaceable.
Plus self-hosted and custom endpoints behind the same client contract.
Keys map to tenant, team, and app. Roles narrow what a caller can invoke.
OpenAI-compatible APIs. Pluggable DLP detectors, policy engines, telemetry sinks, and storage.
A briefing map, not a certification. Shared means you still own identity, legal basis, and BAAs.
| Obligation | Framework | Assert control | Evidence | Posture |
|---|
Prompts, policies, and telemetry stay inside the perimeter. No call home is required to route a model you host.
Gateway, config plane, and audit store install in your VPC or a disconnected enclave. Outbound provider access is a choice.
Airgapped tenants route only to models you run until a policy explicitly allows a commercial provider.
Policy packs and software arrive as artifacts you import. Enforcement does not need a license callback.
Packages, not a public rate card. A briefing prices the gate against your volume. The calculator is illustrative.
OpenAI-compatible API, provider routing, keys, and basic attribution for one organization.
Talk to salesTenants, SSO, DLP, budgets, approvals, hot and cold telemetry, staged rollout.
Talk to salesDisconnected install, signed updates, self-hosted model routing, no phone-home.
Talk to salesNot a forecast. Routing savings still have to pass residency, allow lists, and quality floors.
Start with the question your office will ask.
No. The gateway is OpenAI-compatible. Most apps change a base URL and a key. SDKs, a CLI, dashboards, and webhooks cover the rest.
Yes. Residency and airgap are policies. The router only considers endpoints that satisfy them, including self-hosted models on your network.
Before transmission. Detectors can redact, tokenize, block, or transform. A credential hit can deny the call with no provider request.
The matrix maps obligations to controls and evidence. It is not a certification, a completed audit, or a substitute for your assessors.
Budgets and quotas sit in the policy engine. Over-budget traffic can be denied or shifted to a cheaper allowed model.
Inventory the client, issue a tenant key, point the base URL, compare the audit row.
Allow list, residency, and a credential block. Add budgets after the path is proven.
What stays on the network, which models are local, and how signed bundles enter.
Assert is the control plane for model traffic. This page is a product sketch for a briefing, not a customer list.
Sit on the prompt. Inspect it, route it, record it. Do not make the application learn each provider.
No invented customer logos, no certification badges, no outcome metrics we have not measured with you.
Bring one app, one data class, and last quarter’s model invoice. Leave with a policy and a payback sketch.
Developer on the client. CTO on the path. CIO on tenants. CISO on what never leaves. CFO on who pays.