Amazi AI Gateway

What the gateway does

Amazi AI Gateway gives applications in a project managed access to OpenAI, Anthropic and DeepSeek. It is separate from the external Integrations catalog: provider connections have their own project tab, plan capability, credentials, model catalog, and usage accounting. This feature controls AI services used by your applications; it does not enable or disable the Amazi agent in project chat. Chat work still requires the applicable project permissions and credits.

Open Gateway in the project and connect it when your role and the project owner's plan allow the action. Amazi enables the available providers, creates and stores their managed credentials, and shows their environment names. Select a provider to inspect its credentials, endpoints and models. Disconnecting removes Gateway access and its voice-agent routes without turning providers into normal integrations.

The Providers section lists OpenAI, Anthropic and DeepSeek even before you connect Gateway. Connecting enables all available providers together; each provider has its own managed API key and base URL. If Gateway is already connected and more providers become available, choose Enable providers to add them while keeping your existing provider credentials.

Use the centered tabs below the header: Providers opens first and contains provider credentials and model catalogs. Expand Available endpoints above Available models to see the selected provider's supported HTTP paths and how to append them to its base URL. Expand Available models to see a table with model names, API identifiers, categories and descriptions. Each tab has one information block. SIP Realtime contains OpenAI voice agents, their Realtime addresses and SIP details. Switching tabs preserves your selected provider and unfinished voice-agent form.

If Gateway cannot load, use Retry. A loading error does not mean Gateway has been disabled for your account; plan and role restrictions have separate notices.

Use it from server code

Gateway credentials are server secrets:

• call the provider from a backend or service application;
• use the exact environment names shown by Amazi;
• never copy a credential into chat, source control, a browser variable, or frontend code.

The agent receives instructions and variable names, not the saved secret value.

Ordinary HTTP requests

In backend code, send provider-compatible requests to the injected base URL and authenticate with the injected key as a bearer token. Select a model from the provider catalog shown by Amazi. It lists model names, exact API identifiers, categories and descriptions. The available providers and models table below is loaded from Amazi's current catalog. The agent can inspect endpoint, format and pricing metadata when selecting a model for your application.

Anthropic uses its native Messages format; DeepSeek uses Chat Completions. For both, JSON responses and streaming are supported. Their Gateway connections provide text generation, reasoning and tool calls. Check the chosen model's input formats before attaching images.

Each prefix supplies _BASE_URL and _API_KEY. With the Anthropic SDK, use the injected base URL and key directly: the SDK adds /v1/messages and sends x-api-key. Direct HTTP requests can use that header or Authorization: Bearer <key>; include anthropic-version: 2023-06-01 and max_tokens in the request where required. With an OpenAI-compatible DeepSeek client, set its base URL and API key to the DeepSeek variables. Do not add a second /v1 or send a provider's request body to a different provider's URL.

For conversations with tools, preserve the provider's returned assistant messages and matching tool results. This includes Anthropic thinking signatures and DeepSeek reasoning_content; dropping them can make the next request fail.

These connections support your application's own function tools. Provider-hosted search, code execution, premium speed or regional processing options are unavailable through these routes. Anthropic's standard five-minute and one-hour prompt caches are supported and metered separately.

The operation paths below are defined by the OpenAI API. Amazi AI Gateway forwards supported requests to the corresponding OpenAI endpoints using those same paths. The OpenAI base URL ends in /openai/v1. Append the operation path from this table to it; do not append another /v1. Every operation below uses POST:

For example, transcription uses <base URL>/audio/transcriptions. Sending to the base /v1 alone does not select an operation, and /transcribe is not a supported path. Choose an exact model identifier that supports the operation and input format. A model's category does not choose the route automatically, and Gateway does not expose every provider endpoint.

For JSON operations, set Content-Type: application/json and send the provider's JSON body. For example, POST <base URL>/embeddings accepts {"model":"<embedding-model-from-the-catalog>","input":"Example text"} with the bearer key.

For audio transcription, send file and one text model field as multipart/form-data. The fields may appear in any order. With FormData, let the HTTP client generate the Content-Type header and boundary. In this backend example, modelName is the selected transcription model and audioBlob contains the audio file:

const baseUrl = process.env.AMAZI_AI_GATEWAY_OPENAI_BASE_URL;
const key = process.env.AMAZI_AI_GATEWAY_OPENAI_API_KEY;
const form = new FormData();
form.append("file", audioBlob, "recording.wav");
form.append("model", modelName);
form.append("response_format", "json");
const response = await fetch(baseUrl + "/audio/transcriptions", {
  method: "POST",
  headers: { Authorization: "Bearer " + key },
  body: form,
});
if (!response.ok) throw new Error("Transcription failed: " + response.status);
const transcript = await response.json();

Set the request timeout in your application's HTTP client or provider SDK. Gateway does not impose a model-execution deadline. Your application can cancel a request by closing its connection; otherwise it waits for the provider's result or error.

For OpenAI Responses, Anthropic Messages and DeepSeek Chat Completions streams, your backend can request live progress with the X-Amazi-Stream-Mode: live header and stream: true. Text and reasoning summaries arrive while the model works. A response marked x-amazi-live-stream: 1 reports its settled cost in an amazi.billing event containing chargedCredits and requestId, before the completion event. Native provider events keep their original format. Count this charge once per request ID; if it is missing, show the cost as unavailable.

Realtime WebSocket

OpenAI Realtime is available to server applications. Convert the injected OpenAI base URL from https to wss (or http to ws), append /realtime?model=<exact-model-name>, and send the same bearer credential in the WebSocket upgrade headers. Realtime events retain the OpenAI protocol. Do not request a browser client secret and do not put this server credential into frontend code.

Amazi does not impose a fixed call duration. A session remains connected while the provider is available and the project owner has credits. Completed responses are metered during the session. An Amazi transport failure arrives as an amazi.gateway.error event containing an actionable code, message, requestId, stage, and retryable flag; use these fields in safe backend logs.

Phone voice agents

After Gateway is connected, open SIP Realtime and create one or more OpenAI voice agents. Each agent has an optional name, optional backend configuration and lifecycle webhook URLs, and its own SIP address. Copy that address to the forwarding destination at any compatible SIP operator. Deleting an agent revokes its SIP address without changing other agents. Each card labels the saved configuration endpoint, call webhook, and Realtime WebSocket URLs so you can distinguish their purposes without opening the edit form. If creation reports that the SIP service is unavailable, the agent was not saved. Retry later or contact Support if the problem persists.

Without a configuration URL, the agent uses the default session settings and the first Realtime model in the current catalog. Add a configuration URL for your own instructions, audio settings or tools. When one is set, Amazi sends an authenticated GET to it for each session. A preview backend may be used only for a manual check while its environment is running; an active phone route must use a production application URL so preview sleep cannot replace the JSON response with a wake response. The URL must use HTTPS, be reachable from the public internet, and must not rely on browser cookies or interactive user authentication. Verify its bearer token against the server-only AMAZI_VOICE_CONFIGURATION_SECRET environment variable and return a provider-native Realtime session object containing model. Request headers also identify the session, voice agent, provider, mono PCM16 24 kHz audio contract, and event WebSocket URL. Project backend code can connect to that URL with the selected provider's injected Gateway key to receive provider events and return tool results unchanged. This keeps business tools and database access in the project; Amazi transports the events but does not execute project functions.

Credits and usage

In project chat, type / and choose Check project usage to open the account-wide AI Gateway page with this project selected. The filter stays selected even when the project has no usage yet; you can choose another project or all projects there.

Archived and permanently deleted projects remain in usage statistics with separate Archived and Deleted labels. You can still select them to inspect past usage, and their charges remain included in the totals. Restoring an archived project removes its archive label.

To display the exact cost inside your application, read the x-amazi-charged-credits and x-amazi-gateway-request-id headers from successful HTTP responses on your server. Sum the integer credit amounts once per unique request ID. This uses Amazi's settled charge, including cached tokens and rounding. If the headers are absent, show the cost as unavailable rather than zero.

Gateway requests from a project are charged to the current project owner's Amazi credit balance. They share the account's plan and purchased credits with personal agent work. The account-wide Gateway page identifies whether usage came from an application gateway or the project agent. Usage distinguishes project-agent, image-generation and audio-transcription operations; older Research and Code records remain in history, alongside provider, model, project, status, timing, token totals, and credit usage.

If the owner's plan no longer includes gateway access, existing connection records remain but provider use stops until eligibility returns. Contact Support for a provider-side balance or configuration problem; manage your account credits on Plan.

Available providers and models

OpenAI

Anthropic

DeepSeek