How to Find Ollama, LM Studio, and llama.cpp on Employee Laptops

Signals each runtime leaves

Ports are configurable and defaults change between versions, so match on process names first and use ports as a second signal.

Others include Jan, GPT4All, LocalAI, vLLM, and Docker's model runner. Check their process names rather than relying on a port list.

The tools named here are examples, not endorsements. Use the ones you already run.

Device management inventory

Use your device management tool's installed-software report, in Jamf, Microsoft Intune, Kandji, or similar, to find the Ollama and LM Studio apps. Add a custom inventory script that runs the checks above and reports runtime names, versions, and model names only.

osquery

The processes, listening_ports, apps or programs, homebrew_packages, and python_packages tables cover the same signals. A starter query:

SELECT p.name, p.path, l.address, l.port
FROM listening_ports l JOIN processes p USING (pid)
WHERE p.name LIKE '%ollama%' OR p.name LIKE 'llama-%'
   OR l.port IN (11434, 1234, 9931);

Leave out the cmdline column: command lines can contain secrets, such as a key passed to llama-server --api-key. Elastic's osquery integration ships prebuilt packs that inventory local LLM runtimes, AI coding agents, and MCP servers.

EDR process telemetry

Process-start and file-creation events catch runtimes starting and model files arriving. Splunk publishes detections for local LLM frameworks built on this kind of telemetry.

Network signals, with limits

DNS or proxy hits on model registries, such as the Ollama library or Hugging Face, show downloads, not use. Treat them as a lead to follow up on the machine.

Behest Radar recognizes local models run with Ollama, LM Studio, or llama.cpp, and names the model. It recognizes them by the shape of the request, not by port or process name, so a model served on a non-default port still shows up as a local model with its host; it does not show which runtime is serving it. It also keeps a local list of suspects: unknown hosts whose traffic looks like AI inference, for a person to confirm.

It is observe-only and does not change, slow, or stop anything. A standalone install makes no network connections of its own. It does not replace an installed-software inventory: it shows AI in use, not every file on disk. Check the download page for the installers available today.

Frequently asked questions

Can network monitoring detect local LLMs?

Only partly. DNS, proxy, and CASB logs can show a model being downloaded, for example from Hugging Face or the Ollama library, but once the model is on the laptop it answers without network traffic. To see local LLMs in use, check the machine: running processes, listening ports, model files, and installed apps.

What port does Ollama use?

Ollama listens on port 11434 and binds to 127.0.0.1 by default, so only the same machine can reach it. The OLLAMA_HOST setting changes that address. If you find Ollama listening on 0.0.0.0 or a network interface, other machines may be able to use it without authentication, so treat it as an exposure to fix.

Are local LLMs a security risk?

They carry different risks from cloud AI. Prompts stay on the machine, which helps keep data inside the company. The risks are unvetted model files and licenses, local AI servers exposed to the network without authentication, and outdated runtimes with known vulnerabilities. Most organizations do better sanctioning a local AI setup than banning it.

How do I find local LLMs on every company laptop?

Use tools you already have: your device management tool's installed-software inventory finds the apps, osquery finds runtime processes and listening ports, and EDR telemetry shows runtimes starting. Behest Radar recognizes local models run with Ollama, LM Studio, or llama.cpp. Machines connected to Behest Control appear in one device inventory, where local model use shows up without the model name, which stays on the machine.

Does Behest Radar see local models?

Yes. Behest Radar recognizes local models run with Ollama, LM Studio, or llama.cpp by the shape of the request, on any port, and names the model, though not which runtime serves it. It is observe-only and does not change, slow, or stop anything. Radar does not read prompts by default; only Behest Control's full sharing level adds a redacted prompt excerpt.

Do local LLMs cost anything?

There is no per-token bill, so local models never appear on an AI provider invoice or in expense reports. The costs are hardware, power, and the time people spend running them, plus the risk of using a model whose license does not allow commercial use. Record them in your AI inventory like any other tool.