> ## Documentation Index
> Fetch the complete documentation index at: https://docs.korve.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Gateway

> Provider-neutral model routing with hard budgets, allowlists, retries, and redacted audit events.

AI Gateway gives deployed server code one OpenAI-compatible chat-completions endpoint. Applications
select stable Korve model aliases while the platform owns routing, retries, usage accounting, and
credential isolation.

## Declare a gateway

```toml theme={null}
[[projects.aiGateways]]
name = "primary-ai"
environment = "production"
allowedModels = ["korve/fast", "korve/balanced"]
monthlyBudgetUsd = 25
timeoutMs = 30000
maxRetries = 1
redactPrompts = true
```

The supported aliases are `korve/fast`, `korve/balanced`, and `korve/reasoning`. A gateway must
allow at least one alias. Monthly budgets range from $1 to $100,000, timeouts from 1–120 seconds,
and automatic retries from 0–3.

```bash theme={null}
korve manifest plan -f korve.toml
korve manifest apply -f korve.toml
```

## Complete from a deployed server

```ts theme={null}
import { createKorve } from "@korve-dev/sdk";

const result = await createKorve()
  .ai("primary-ai")
  .complete({
    model: "korve/fast",
    messages: [{ role: "user", content: "Summarize this support ticket." }],
    maxTokens: 600,
  });
```

Korve injects the project-and-environment runtime capability. Never copy it, or a gateway key, into
a browser bundle. A direct server client can call:

```text theme={null}
POST /v1/runtime/ai/primary-ai/chat/completions
Authorization: Bearer $KORVE_RUNTIME_TOKEN
Content-Type: application/json
```

The body supports `model`, `messages`, `stream`, `max_tokens`, and `temperature`. Streaming uses
`text/event-stream`. Requests can contain at most 256 messages and 1,048,576 encoded bytes.

## Security, privacy, and spend

* Requests for models outside the gateway allowlist fail before inference.
* The hard monthly gateway budget stops new requests with `429` when exhausted.
* Prompt bodies are excluded from audit events by default. With `redactPrompts = false`, audit
  storage still retains only a redacted, bounded preview—not a complete prompt.
* Gateway key secrets are shown once; listings return metadata only. Revoke a key immediately if
  its one-time value may have escaped.
* Errors are provider-neutral and never include upstream response bodies, identifiers, or
  credentials.
* Deleting a gateway permanently revokes all of its gateway keys.

Gateway creation, configuration changes, key creation, and inference require active billing.
Inference is metered against the public `ai.gateway.inference` rate-card entry and the gateway's own
hard budget. Inspect current-month tokens, spend, and remaining budget with
`aiGateways.usage`.

See the complete [AI Gateway API](/api-reference/aiGateways).