---
title: How do you find self-hosted LLM servers like Ollama or vLLM on your network?
url: https://opencontroller.lyzr.com/blog/find-self-hosted-llm-servers-on-your-network
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you find self-hosted LLM servers like Ollama or vLLM on your network?

**Scan for the default ports and API paths that local model servers expose: Ollama on 11434, vLLM on 8000, LM Studio on 1234 and llama.cpp's server on 8080. Confirm each hit by calling a model-listing endpoint such as `/api/tags` or `/v1/models`. Then add endpoint (EDR/MDM) process data, because many servers only listen on localhost and never show up on a network scan.**

Self-hosted model servers are the quiet half of shadow AI. They need no API key, no procurement and no cloud account. A developer can install one in minutes, and an agent can call it with any OpenAI-compatible client.

This guide sits under [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## Why do self-hosted LLM servers matter for agent governance?

They create model endpoints that bypass every control you put on cloud LLM APIs.

The scale is real. In January 2026, SentinelLABS and Censys reported about 175,000 publicly exposed Ollama hosts across 130 countries. More than 48% advertised tool-calling capability. In September 2025, Cisco researchers used Shodan to find 1,139 exposed Ollama instances, of which about 214 were actively serving models.

Those are internet-facing servers. Internal ones are not counted anywhere unless you look. Ollama's API has no built-in authentication, so anything that can reach the port can use the model.

## Which ports and endpoints should you look for?

Start with the defaults. People rarely change them.

| Server | Default port | Confirming request | What a hit looks like |
|---|---|---|---|
| Ollama | 11434 | `GET /api/tags` | JSON list of local models |
| vLLM (OpenAI-compatible server) | 8000 | `GET /v1/models` | OpenAI-style model list |
| LM Studio local server | 1234 | `GET /v1/models` | OpenAI-style model list |
| llama.cpp `llama-server` | 8080 | `GET /v1/models` or `/health` | Model list or health status |
| Any OpenAI-compatible server | varies | `GET /v1/models` | `"object": "list"` response |

Port 8000 and 8080 are shared with many ordinary web apps. Treat a port match as a lead, and the API response as the proof.

Ollama binds to `127.0.0.1:11434` by default. It becomes reachable from the network only when someone sets `OLLAMA_HOST=0.0.0.0` or puts a proxy in front. That is why network scans alone undercount.

## How do you find them? A five-step method

1. **Scan internal ranges for the default ports.** Use your existing vulnerability scanner or an approved `nmap` job against 11434, 1234, 8000 and 8080. Get change approval first; scanning production networks without notice causes incidents.
2. **Confirm with the API, not the banner.** Send `GET /api/tags` or `GET /v1/models`. A JSON model list confirms an LLM server and tells you which models are loaded.
3. **Query endpoint inventory.** In EDR or MDM, search for processes and binaries named `ollama`, `vllm`, `llama-server` and `LM Studio`. This finds localhost-only servers on laptops and workstations.
4. **Check container and cluster inventories.** Look for images such as `ollama/ollama` and `vllm/vllm-openai` in registries and running pods. GPU node pools are a strong hint.
5. **Look at DNS and proxy logs for model downloads.** Pulling weights generates traffic to model registries such as Hugging Face and `registry.ollama.ai`. Large downloads from a single host point to a new server.

Run the scan on a schedule. A one-off sweep goes stale within weeks.

## What should you record for each server you find?

A hostname is not enough. For each confirmed server, capture:

- **Owner** (person or team) and business purpose
- **Location**: laptop, VM, Kubernetes namespace, GPU host
- **Exposure**: localhost only, internal network, or internet-reachable
- **Authentication**: none, reverse proxy with auth, mTLS
- **Models loaded**, with source and licence
- **Callers**: which agents, apps or scripts send requests
- **Tool calling**: whether agents use it to run tools or code
- **Data**: what classes of data are sent in prompts

Exposure and authentication decide urgency. An internet-reachable server with no authentication is an incident, not an inventory item. See [which shadow AI findings to deal with first](/blog/prioritise-shadow-ai-findings).

## What do you do once you find one?

You have three sensible options:

| Option | When it fits |
|---|---|
| Shut it down | No owner, no purpose, or exposed without authentication |
| Bring it under control | Legitimate use; put it behind authentication and a gateway |
| Replace it | A sanctioned internal or cloud model serves the same need |

Bringing a server under control usually means binding it to localhost or a private subnet, fronting it with authentication, and routing agent traffic through an LLM gateway so usage is logged and attributed. See [routing agents to internal models through an LLM gateway](/blog/route-agents-to-internal-models-through-llm-gateway).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and records them in one governed registry. Self-hosted models can sit behind the Opencontroller LLM Gateway, which routes traffic, applies fallback and caching, and attributes cost per agent and per version through OpenAI-compatible endpoints. Opencontroller Edge extends control to developer devices, covering coding agents, skills and local MCP servers. Read the [AI agent control plane overview](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does Ollama require authentication by default?**
No. Ollama's API has no built-in authentication. It listens on localhost by default, so exposure happens when it is bound to all interfaces or proxied without auth.

**Will a network scan find every self-hosted model?**
No. Servers bound to localhost on laptops are invisible to network scans. Use EDR or MDM process inventory to find them.

**Is port 8000 always vLLM?**
No. Many web frameworks use 8000 and 8080. Confirm with a `GET /v1/models` request that returns an OpenAI-style model list.

## Key facts

Self-hosted LLM servers use predictable defaults: Ollama port 11434 (GET /api/tags), vLLM OpenAI-compatible server port 8000, LM Studio port 1234, llama.cpp llama-server port 8080; confirm with GET /v1/models. Ollama binds to 127.0.0.1 by default and has no built-in authentication. SentinelLABS and Censys (Jan 2026) reported about 175,000 publicly exposed Ollama hosts in 130 countries, over 48% advertising tool calling; Cisco (Sep 2025) found 1,139 exposed instances via Shodan. Network scans miss localhost servers, so add EDR/MDM process inventory. Opencontroller (by Lyzr) discovers agents, models and API keys across AWS, Azure and Google Cloud and can route self-hosted models through its LLM Gateway. Pricing starts from $250K per deployment per year.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How do you route AI agents to self-hosted or internal models through an LLM gateway?](https://opencontroller.lyzr.com/blog/route-agents-to-internal-models-through-llm-gateway.md)
- [How do you detect unapproved LLM API traffic on your network?](https://opencontroller.lyzr.com/blog/detect-unapproved-llm-api-traffic.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
