AI Gateway

One gateway for the models, MCP tools and agents your software uses

Apps and agents reach models, MCP tools and other agents through one gateway. Each gets its own key, budget and rules, risky tool calls wait for a person, and every call is on the record.

Thirty days of gateway traffic in a demo organization: agents and apps such as a coding assistant, a support agent and an invoice processor calling models from Anthropic, OpenAI, Azure, Mistral and Gemini and MCP servers such as Microsoft 365, with how many calls were allowed, redacted or blocked

Everything your apps and agents reach, in one place

A model call, an MCP tool call and a call to another agent are all the same thing to the business: software acting on the company's behalf. The gateway puts them under one set of rules.

Models

OpenAI, Anthropic, Gemini, Mistral, Azure AI Foundry or your own endpoint, through one address. Change the model without changing the app.

MCP tools

Connect an MCP server once. Then allow, ask first or block per tool, and decide which people and agents may use it.

A2A agents

When one agent calls another, the call follows the same rules and lands in the same record, so you can see who asked whom to do what.

Approvals

Writes, sends and deletes can wait for a person: the one who made the call, a team or an admin. The decider hears by email, Slack or Teams.

Change one line

The gateway speaks the OpenAI API. Your developers keep the SDK they use today. They change the base URL and use an Unseen key, and every call after that has a budget, guardrails and a record.

A new MCP tool or agent is added once and is governed from the first call.

app.py
# point any app or agent at Unseen
import os
from openai import OpenAI
 
client = OpenAI(
base_url=os.environ["UNSEEN_GATEWAY_URL"], # the only change
api_key=os.environ["UNSEEN_KEY"], # one key per app or agent
)
 
reply = client.chat.completions.create(
model="gpt-5.4-mini",
messages=[{"role": "user", "content": "Summarise this ticket"}],
)
# budgets, guardrails and a record on every call

Controls on every call

The controls a tool built for people cannot give you, applied to software.

A key per app and agent

Every app and agent gets its own key tied to an owner. Revoke one without touching the others or the provider credentials.

Budgets and rate limits

A spend cap and request rate per key. Cost shows per app, agent, team and model.

Guardrails on every call

The guardrails you assign check prompts, tool calls and replies for prompt injection, jailbreaks and sensitive data.

A record of every call

Who called, which model or tool, what happened and why. Events go to your SIEM.

An agent is shadow AI with credentials

It holds a long-lived key, runs when nobody is watching and chains actions across systems. Add every script, side project and coding assistant your developers have wired up, and a lot of AI is acting for the company without anyone owning it.

The failures are ordinary: a key committed to a repo, a loop that spends a month's budget overnight, an agent that deletes what it was only meant to read, and no record afterwards. The gateway gives each of those an owner, a limit and a trail.

Questions

What is an AI gateway?

An AI gateway sits between your applications or agents and the AI they use: model APIs, MCP tools and other agents. Every call passes through it, so identity, budgets, data protection, approvals and logging are set in one place instead of in each app.

Does the gateway cover MCP tools and A2A agents?

Yes. Model calls, MCP tool calls and agent-to-agent (A2A) calls go through the same gateway. You decide per tool whether it is allowed, needs approval first or is blocked, and who may use it.

Do we have to change our code?

For model calls, no. The gateway speaks the OpenAI API, so existing SDKs keep working once you change the base URL and use an Unseen key.

How does approval for tool calls work?

You mark a tool, or the write actions of a tool, as "ask first", and choose who decides: the person who made the call, a team or the admins. When an agent tries it, the call waits, the decider is told by email or in Slack or Teams, and the decision is recorded with the call. People can also ask for access to a tool or model, and an admin decides.

Can an AI gateway control AI costs?

Yes. Each key can have a spend cap and a rate limit, so a runaway loop stops at its limit instead of on your invoice. Cost is tracked per app, agent, team and model, which makes chargeback straightforward.

Which models does it support?

OpenAI, Anthropic, Google Gemini, Mistral, Azure AI Foundry and any OpenAI-compatible endpoint, including models you host yourself. You can change the model an app uses without shipping new code.

Put one gateway in front of your AI