What is an AI gateway?
Reviewed October 2026
An AI gateway is a single path that applications, coding tools and agents use to reach AI models. Every call that crosses it is checked for identity, held to your policy and spend limits, routed to an approved model, and recorded. It is also called an LLM gateway. An LLM router is the routing part on its own; a gateway adds the identity, policy and record around it.
What an AI gateway does
A gateway puts one control point in front of every model your organization uses, so the rules are applied once rather than inside every application. On that path it does six things:
- Identity: establishes which person, team or agent is calling, with which rights
- Policy enforcement: allows only approved models and tools, and applies guardrails such as blocking or masking sensitive text
- Spend control: request, token and dollar limits by team, project or agent
- Routing: one endpoint in front of several model providers, with fallback when one fails
- Audit trail: caller, model, tokens, cost and the policy decision for each call
- Tool calls: the same identity and policy for MCP tool servers and business APIs, not only for model calls
AI gateway vs API gateway
An API gateway manages HTTP APIs: authentication, rate limits per request, and routing to services. An AI gateway is built for model traffic. It counts tokens and cost rather than requests, knows which model and provider a call is going to, holds the provider keys so applications and agents do not, applies guardrails to prompts and responses, and keeps a record that a risk or audit team can read. Many AI gateways reuse API gateway patterns underneath.
AI gateway vs LLM router
An LLM router decides which model or provider a request goes to, often by cost, latency or availability, and retries on failure. That is one function of an AI gateway. A router alone does not establish who is calling, decide whether they are allowed, enforce spend limits or keep a record of the decision. If you need to answer who used which model, for what, under which rules, you need the gateway around the router.
Why regulated enterprises use one
A control inside a thousand applications is a thousand implementations, a thousand reviews and a thousand places it can be wrong. At the gateway it is one. Regulated institutions use an AI gateway so that model access follows a written policy, provider keys stay out of applications and agents, spend is capped before the bill arrives, and every call leaves a record in storage the institution owns.
Ways to run an AI gateway
Gateways differ less in features than in who runs them, where they run and where the record lands. Most options fall into four approaches: a vendor-hosted shared service, a cloud provider's gateway, open source you run yourself, or a managed, dedicated gateway run for you on your own keys. The enterprise LLM gateway comparison sets out the criteria for choosing between them.
Where Reign Gateway fits
Reign Gateway is a managed AI gateway: identity, policy and spend limits on every model call, with a record you own, run as a dedicated single-tenant instance. Developers point their SDKs at one endpoint: OpenAI-compatible, Anthropic-compatible or MCP. It works with Claude Code, GitHub Copilot and Cursor, and with apps built on the OpenAI or Anthropic SDKs. Bring your own models, including open-weights models on your own endpoint. It runs on Reign Dedicated or in your own AWS or Azure account. Air-gapped and sovereign deployment are supported. Reign runs, patches and upgrades it 24×7.
See Reign Gateway on one traffic path
Reign Gateway applies identity, policy and spend limits to the AI traffic sent through it and records each call in storage you own.