AI gateway, LLM router, MCP gateway, AI firewall: what is the difference?

A router picks the model, an AI gateway controls who may use it, a firewall reads the words, and an MCP gateway decides each tool call before it runs. Who needs which, and how Verosek does them on one key.

Vats Shah6 min read
[ In short ]

An LLM router picks the model, an AI gateway also controls who may call it and what they spend, an AI firewall reads the words, and an MCP gateway decides each tool call before it runs. Verosek is one gateway for model calls and tool calls: it routes model calls, sets budgets and rate limits per key, decides each tool call against that key's rules and seals a record of every request. Turn on Shield and it scans what goes in and out too, with no second vendor to buy.

[ On this page ]
NameWhat it seesWhat it decidesWho needs it
NameLLM router, or AI routerWhat it seesThe model asked for, its price, its errors and its speedWhat it decidesWhich model or provider answersWho needs itA team that wants several models behind one API, with a fallback
NameModel marketplace, such as OpenRouterWhat it seesThe model asked for, and each provider's price, speed and uptimeWhat it decidesWhich provider answers, with a fallbackWho needs itA team that wants many models on one account and one bill
NameLLM proxyWhat it seesThe call and the key behind itWhat it decidesWhere to forward it and what to logWho needs itOne app that wants one address and its provider keys out of the code
NameAI gateway, or LLM gatewayWhat it seesWho is calling, the model, the tokens and the spendWhat it decidesWho may use which model, within what budgetWho needs itSeveral teams sharing model access
NameAI firewallWhat it seesThe words going in and coming outWhat it decidesWhether the words look like an attack or a leakWho needs itA chat product that strangers can type into
NameMCP gatewayWhat it seesThe tool, its arguments and the agent's keyWhat it decidesWhether this tool call may runWho needs itAgents that reach systems they can change
Table 1. The names, in plain words. Products mix these jobs, so read what a product sees, not what it is called.
[ 01 ]

AI router vs AI gateway: what is an AI gateway for?

An AI router chooses which model answers, and an AI gateway also controls who may call it. A router weighs price, speed and errors, then sends each call to a model that fits, with a fallback when one fails. An AI gateway usually does that too, and adds a key per team, budgets, rate limits and one log of usage. DigitalOcean's guide to LLM routers calls a gateway "a broader control layer", and says the two terms "can refer to similar concepts". You feel the difference the day a second team arrives and finance asks who spent what.

[ 02 ]

What is an LLM proxy?

An LLM proxy forwards your calls through one address and keeps your provider keys out of the code. It logs what passes and decides little else. The words overlap, and many products answer to proxy, router and gateway at once.

[ 03 ]

Is an LLM gateway the same as a model marketplace?

A model marketplace is a router someone else runs: one account, many models, one bill. You get breadth with no setup, and your prompts pass through a third party. An LLM gateway you run yourself keeps your own provider keys and your own traffic, and you do the upkeep. People weighing the two mostly care about catalogue and price, which is a fair way to choose a router.

[ 04 ]

What is an AI firewall, and is it AI security?

An AI firewall reads the words going into and out of a model, and stops what looks like an attack or a leak. You want one when strangers can type whatever they like into your product. It is one part of AI security, and it reads words: a harmful action can arrive in polite ones, and the firewall has nothing to object to. OWASP, the security group, puts prompt injection first on its list of risks for LLM applications (OWASP LLM01), and says "it is unclear if there are fool-proof methods of prevention".

[ 05 ]

What is an MCP gateway?

An MCP gateway sits between your agents and their tools, and decides each tool call before it runs. MCP, the Model Context Protocol, is a standard way for an agent to call tools: its specification calls it "an open protocol" between "LLM applications and external data sources and tools". The gateway reads the call itself: which tool, which arguments, under which agent's key. A $240 refund against a $100 limit is refused, however politely the model asked.

[ One request ]
The modelThe cheaper of two models
The keyThe support agent's, within its budget
The words“Please refund the customer, as agreed.”
The tool callRefund $240, on a key whose limit is $100
Blocked

$240 is over this key's $100 limit, so your system never receives the call.

Figure 1. One request, seen by each kind of product. Pick one. Illustrative values.
[ 06 ]

Who needs only a router?

If your app calls models and nothing it does changes your systems, a router is enough. Several teams sharing a budget need an AI gateway. A chat product open to strangers needs a firewall or a guardrail. Reach for an MCP gateway when an agent can issue a refund, write to a database or send a message. If you need more than one of these, they can be one product: Verosek does them on one key, with Shield on to read the words.

Can it change things?refunds, writes, messages
Can strangers type in?a public chat product
Do teams share access?keys, budgets, one log
Several models, one API?with a fallback
MCP gatewaywhere Verosek sits
AI firewallor a guardrail
AI gatewayor LLM gateway
LLM routerno: the provider's API
nononoyesyesyesyes
Can it change things?refunds, writes, messages
Can strangers type in?a public chat product
Do teams share access?keys, budgets, one log
Several models, one API?with a fallback
MCP gatewaywhere Verosek sits
AI firewallor a guardrail
AI gatewayor LLM gateway
LLM routerno: the provider's API
Figure 2. Four questions that settle which one you need. Illustrative.
[ 07 ]

Does a router give you agent security?

A plain router carries an agent's tool call back to your app without deciding anything about it. The model's reply holds the call, the router returns the reply, and your code runs the refund. Agent security starts when something outside the model reads that call and can say no, as How do you secure AI agents? explains.

The modelasks for a refund
A plain routerreturns the reply
Your appruns the call
Stripethe refund runs
The modelasks for a refund
An MCP gatewayreads the call
Stripenever receives it
RefusedOver this key’s $100 limit
The modelasks for a refund
A plain routerreturns the reply
Your appruns the call
Stripethe refund runs
The modelasks for a refund
An MCP gatewayrefuses: over the $100 limit
Stripenever receives it
Figure 3. The same tool call, with a plain router and with an MCP gateway in the path. Illustrative values.
[ 08 ]

Where does Verosek sit?

Verosek sits in the path of the model call and the tool call, on one key per agent. It routes model calls to your providers with fallback for chat, holds each key to its allowed models, budget and rate limits, decides each tool call against that key's rules before it runs, and seals one record of both. Versioned agents get their own managed key and limits on loops and time. Shield's checks ship in the gateway: turn them on and the same key scans the words too, with no second key and no second record. A router sorts by catalogue and price per token; Verosek decides what each call may do.

If your agents can write to something you care about, book a call and we will look at which of these you need.

[ 09 ]

Frequently asked questions

Is Verosek a router?

It routes model calls, with fallback for chat, and it is a full gateway around them: budgets and rate limits per key, a decision on each tool call and one sealed record.

Is an LLM gateway the same as OpenRouter?

No. People use OpenRouter for routing, and it does that job. An LLM gateway you run keeps your provider keys and your traffic on your side. Verosek adds the decision on each tool call.

Is an AI firewall the same as an MCP gateway?

No. A firewall reads words, and an MCP gateway reads calls. A polite request to refund the wrong customer can pass the first and stop at the second.

Do I need a router and an MCP gateway?

You need each job you actually have. Verosek does routing, tool decisions and, with Shield on, scanning in one request, so one product covers all three.

Is an open-source router enough?

For routing, often yes. A plain router does not decide what your agent's tool calls may do, because that is not its job. Some projects add it, so check what yours reads.

Keep reading

[ Field guide ]

What is prompt injection, and what actually stops it?

Prompt injection is text that makes a model follow the attacker. No filter catches all of it, so the control that holds sits outside the model: a rule on the tool call. Verosek decides that call before it runs.

6 Nov 20264 min read
[ Benchmark ]

Eight AI model setups, 19 MCP attack tasks: how often did the model try to leak a planted secret?

Self-reported: a lower bound of 22.7 to 61.0, and 9 to 18 of 19 tasks fired, on one run per route. Assay's results, limits kept. They rank these configurations, not the models, and say nothing about any gateway.

6 Nov 20266 min read
[ Field guide ]

How do you restrict what SQL an AI agent can run?

Check each statement against rules set outside the AI. Verosek lets through what the agent's key allows and refuses the rest, with the reason, before the database sees it.

6 Nov 20265 min read

Rules for every tool call Verosek runs. Decided before it executes, outside the model.

Still curious? Ask your favourite AI about Verosek.