From project to receipt
1. Prepare the project
Check the current policy and manage scoped, expiring keys. A secret key is shown once; the server keeps its hash, not a recoverable copy.
2. Authorize the transfer
Review the destination and model in the current notice and explicitly consent before sending text. Configuration changes invalidate previous consent. An integration key does not authorize account administration.
3. Check the same reference
Use a stable idempotency key and check the receipt. If the external outcome is uncertain, the budget reservation remains held and the request is not repeated automatically.
Capabilities confirmed by the service
Local management and external inference have different requirements. Unconfirmed capabilities are shown as unavailable; enabling an option does not establish provider health or an SLA.
Chat Completions: text subset
Checking…Project integration keys
Checking…Request and token budgets
Checking…Local allowlisted metadata queries
Checking…Redacted receipts and metadata
Checking…Read a retained encrypted response
Checking…Tariff policy estimate
Checking…Streaming
Streaming not implementedMultimodal inputs
Multimodal inputs not implementedRemote tools
Remote tools not implementedClient-supplied providers
Client providers are not allowed
A text subset, not full compatibility
The public vb-text alias accepts the documented Chat Completions subset with system, user and assistant text messages. Full OpenAI API compatibility, streaming, remote tools, functions and multimodal inputs are not offered.
The server configures the destination and model. Clients cannot supply provider URLs, upstream credentials or arbitrary model routes. URLs written in a message are treated as text and are not fetched.
Metadata queries use a local allowlist of project, capabilities and usage information. They are not document, transcript or XML connectors and do not execute client functions or funds actions.
Content, keys and retention
Content retention defaults to zero. Only an explicit option, with encryption configured and an authorized scope, permits temporary encrypted response retention. The transfer notice describes the destination; zero retention at an unverified provider is not promised.
Receipts contain redacted metadata and pseudonymous request fingerprints: they are not anonymized data. They do not include plaintext prompts or responses, secret keys or provider credentials.
Deleting content revokes reads and queues a purge. Minimal receipts and pending reservations may remain to reconcile the attempt. Dossier expiry does not by itself release an uncertain reservation or revive the dossier; a new period requires a new project.
Uncertainty without duplicate attempts
If it is unknown whether the provider executed an attempt, the receipt requires reconciliation and freezes the reservation. Check the same reference; there is no automatic replay or regeneration of an absent or expired response.
Unknown usage or cost does not mean zero. Request and token limits are separate from ASR minutes and other pilots. Where a server tariff policy exists, its estimate is not an invoice, a charge or the provider’s billed amount.
If key creation confirmation is lost, a lookup returns only its identifier and prefix, not the secret. Explicitly revoke the unseen key and create another operation; it is not silently regenerated.
Pilot limits and access
AI Gateway requires its own ai-gateway pilot and gateway_request budget, alongside current token limits and consent. It does not include new prices, automatic payments or unlimited usage.
Current limits could not be confirmed. Budgets, prices, retention and inference availability are not assumed; check the project policy in the dashboard.
Each project belongs to one account. Project keys have a limited scope and do not enable other modules. A pending request does not activate the pilot or authorize data transfer.
Request pilot access
Register a Enterprise AI gateway request for your account. This does not create a subscription, activate permissions or execute a payment.