---
header: Fastly AI Runtime Control
lang: en
last_updated: '2026-09-15'
url: https://docs.fastly.com/products/fastly-ai-runtime-control
---

[Fastly AI Runtime Control](https://www.fastly.com/documentation/guides/platform/ai-runtime-control/) (ARC) is a control plane that sits between your applications and large language model (LLM) providers, giving you a single place to manage access, cost, observability, and resilience for your AI traffic. ARC does not host or run LLMs. Instead, it routes requests through a single Fastly-managed endpoint to upstream providers and to your own self-hosted model endpoints, using the provider credentials you supply.

## How it works

Your applications and developers send requests to LLM providers through ARC instead of calling those providers directly. You can also register a self-hosted model endpoint as a provider.

You configure your provider credentials in ARC once, then issue virtual keys that your applications and developers use in place of raw provider keys. A virtual key is an ARC-issued credential that maps to an ordered list of provider and model pairs. If one fails, ARC attempts the next in the list. ARC routes each request to the first available pair, returns the model's response, and records the request for accounting and logging. For each request, ARC captures the virtual key that made the request, the session when the client supplies a session ID, the provider and model, the message sent and the response returned, input and output token counts, and a timestamp.

Because traffic passes through ARC, you can attribute usage to the virtual key that made each request and to the session it belongs to. This gives you per-key and per-session visibility into AI usage.

With ARC you can:

- Route requests to multiple LLM providers and your own self-hosted endpoints through one endpoint, using your own provider accounts and API keys.
- Issue virtual keys that your applications and developers use in place of raw provider keys, and refresh or revoke them at any time.
- Track cost and token usage per virtual key and per session, and search your request records.
- Set model failover so requests continue when a provider or model is unavailable.
- Set rate limits by requests or tokens per minute, set budget limits, and get alerts as spending approaches a threshold you set.
- Add prompt injection detection with [Fastly AI Firewall](https://docs.fastly.com/products/fastly-ai-firewall), a separately purchased add-on security module.

## Prerequisites

To use AI Runtime Control, you must supply your own provider accounts and API keys for each LLM provider you route to.

## Limitations and considerations

Keep in mind the following considerations when using ARC:

- **Upstream providers.** Availability and performance of ARC depend in part on the upstream providers you route to. You remain responsible for your relationships and agreements with those providers, and for the content of the requests and responses that pass through ARC.
- **API key handling.** Your LLM provider authorization headers transit the ARC gateway via TLS and are forwarded to the upstream provider. Keys are never persisted to disk, included in request records, or accessible in Fastly debug tooling.
- **Non-text content is not viewable in logs.** Request records are JSON, so images, video, and other non-text content are recorded as encoded text rather than as viewable files.
- **Data retention.** Fastly captures request data to provide usage accounting and logging. All data is stored in the United States on Fastly's own infrastructure or via our [sub-processors](https://docs.fastly.com/products/sub-processors) for a maximum of 7 days. Data log formats and their views cannot be customized.

## Billing

Fees for ARC are based on the number of requests received for processing by the ARC gateway. Fees for ARC do not include any charges you incur directly from your LLM providers for the underlying model usage.