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

# AI surface

> Find the MCP servers, LLM gateways, model runtimes, and agent endpoints your organisation exposes

AI infrastructure is a new kind of external surface, and it is being deployed
faster than it is being inventoried. An MCP server, an LLM gateway, or a model
runtime is a network service like any other, except that it often ships with
no authentication, reaches internal tools on purpose, and is stood up by a team
that does not think of itself as running production infrastructure.

AI surface finds that infrastructure on your surface and tells you what it is.

## Surface types

| Type             | What it is                                                     | Why it matters                                                         |
| ---------------- | -------------------------------------------------------------- | ---------------------------------------------------------------------- |
| `mcp_server`     | A Model Context Protocol server exposing tools to an AI client | Tools often reach internal systems, and many servers ship without auth |
| `llm_gateway`    | A proxy or router in front of one or more model providers      | Holds provider API keys and can be used to spend them                  |
| `model_runtime`  | A server running a model directly                              | Exposes inference, and sometimes the model itself                      |
| `agent_endpoint` | An endpoint an autonomous agent acts through                   | Actions taken here are taken as you                                    |
| `ai_application` | A user-facing application built on a model                     | Carries the prompt, the data, and the user session                     |

Anything recognised as AI infrastructure but not yet classified is kept as
`unknown` rather than dropped, so the surface stays complete.

## What a row tells you

Each AI surface row identifies both the service and the AI stack behind it:

* **Where it is** - hostname, IP, port, protocol, URL, and HTTP title and status.
* **What it runs** - the provider, the product, and the model name where these
  can be determined.
* **How exposed it is** - a risk level and an exposure summary in plain words.
* **Why we say so** - the underlying evidence, carried with a schema so you can
  parse it rather than scrape it.

## Triage

Rows open filtered to `open` status, so you start on what has not been dealt
with. From there, filter by risk, provider, product, or the source that found
it, and work rows through the same lifecycle as every other finding: accept,
resolve, ignore, false positive, or reopen.

<Tip>
  Start with risk, then provider. A critical row on a provider your organisation
  has never bought is usually shadow AI, and worth a conversation before a
  ticket.
</Tip>

## Hunt AI infrastructure everywhere

AI surface covers the assets in your project. The same classification runs
across the whole internet map, so you can hunt AI infrastructure outside your
own inventory with InfraQL:

```text theme={null}
services.tags.category:software/ai/*
```

Narrow it to a branch of the taxonomy to ask a sharper question:

```text theme={null}
services.tags.category:"software/ai/llm-gateway"
AND country_code:DE
```

<CardGroup cols={2}>
  <Card title="Browse the AI taxonomy" icon="tags" href="/platform/taxonomy">
    See every `software/ai` path, from agents and RAG to model runtimes and
    observability.
  </Card>

  <Card title="Back to the external surface" icon="https://mintcdn.com/infrawatch/gCEz_Bv1hOrMPG8n/images/products/esi.svg?fit=max&auto=format&n=gCEz_Bv1hOrMPG8n&q=85&s=ebc334425db0ee423bdf03e1348a066c" href="/external-surface/overview" width="32" height="32" data-path="images/products/esi.svg">
    Inventory, findings, DNS risks, typosquatting, and secrets.
  </Card>
</CardGroup>