LM Studio headless server: lms, llmster and the local API (Linux, Windows, Mac)

What you get in one minute#

LM Studio can run without its window: a headless background service called llmster, driven by the lms command, serves your local models over an OpenAI-compatible API on port 1234, so your own scripts and apps can use them with no screen and no prompt leaving your machine.

Logo: the official LM Studio brand kit (lmstudio.ai/brand)

LM Studio is a desktop app for downloading and chatting with open models. Its maker also ships llmster, the same engine packaged as a daemon for Linux boxes, GPU rigs and machines where a GUI is not practical. This guide shows the headless route: install llmster, download a model, start the server, check the API, decide who may reach it, and keep it running after a reboot.

At a glance:

• What you install: llmster (the headless daemon) and lms (the command line). The desktop app is optional.
• Where it listens: http://localhost:1234 by default. Other machines reach it only if you bind it to the network on purpose.
• Which API: OpenAI-compatible at /v1, plus LM Studio's native /api/v1 and an Anthropic-compatible /v1/messages.
• Models: downloaded with lms get, loaded with lms load or on demand (JIT). JIT-loaded models unload after 60 minutes idle by default.
• Platforms: Linux, Windows (PowerShell) and Mac (Terminal) all have an official llmster install command. Intel Macs are not supported by the app; llmster on Intel Macs is not documented.
• Not documented: a Docker image, token creation from the command line, TLS on the LM Studio server, llmster-only system requirements.

How it fits together#

Your app sends one request to the port. llmster checks the token if you turned that on, loads the model if it is not in memory, and answers. Everything runs on your computer.

Our diagram, drawn from the official LM Studio docs (server, authentication, JIT loading), read October 2026. Dashed boxes happen only when you turn an option on or the model is not loaded yet.

Before you start#

The hardware row above comes from the desktop app's requirements; the docs publish no separate llmster requirements.

Linux (x64 or ARM64): the desktop app needs Ubuntu 20.04 or newer; the docs call Ubuntu newer than 22 "not well tested". Check your version:

lsb_release -a

Windows (x64 or ARM): the x64 build needs AVX2 in the CPU. Check the Windows version and processor name:

Get-ComputerInfo -Property OsName, OsVersion, CsProcessors

The PowerShell command does not show AVX2; check that on the processor's specification page (our suggestion).

Mac (Apple Silicon): macOS 14.0 or newer and an M1 to M4 chip. Check with sw_vers.

Checklist before you install:

• [ ] Hardware matches the rows above
• [ ] Enough free disk space for the models you plan to download (the docs do not list sizes per model)
• [ ] Port 1234 is free, or you have picked another with --port
• [ ] You know who should reach the API: only this computer (the default), or your network (then turn on authentication first)

Tested on: not tested by DIY Lab Hub. We ran nothing for this page. Commands and values come from the official docs read in October 2026; example outputs are the docs' own and are labelled as such.

Step 1: Install llmster and the lms command#

llmster installs with a one-line script on Linux and Mac, and a PowerShell script on Windows. These are the docs' exact commands.

Linux#

curl -fsSL https://lmstudio.ai/install.sh | bash
lms --help

Our advice: to read the script first, run curl -fsSL https://lmstudio.ai/install.sh -o install.sh, open it, then bash install.sh.

Windows (PowerShell)#

irm https://lmstudio.ai/install.ps1 | iex
lms --help

Our advice: read the script first with irm https://lmstudio.ai/install.ps1 | more. Open a new PowerShell window before you run lms. If lms is not recognised, the official lms README suggests npx lmstudio install-cli.

Mac (Terminal)#

Same commands as Linux:

curl -fsSL https://lmstudio.ai/install.sh | bash
lms --help

Check it worked. The docs show this as example output (your version will be newer):

lms is LM Studio's CLI utility for your models, server, and inference runtime. (v0.0.47)
Usage: lms [options] [command]

Tip: Open a new terminal after the install. Most "command not found" problems are an old window that started before the PATH change.

Step 2: Start the daemon and check it#

lms daemon up starts llmster, and lms daemon status says whether it runs.

Linux#

lms daemon up
lms daemon status --json

Windows (PowerShell)#

lms daemon up
lms daemon status --json

Mac (Terminal)#

Same commands as Linux:

lms daemon up
lms daemon status --json

Check it worked. The docs' example output when running:

{ "status" : "running", "pid" : 12345, "isDaemon" : true }

Tip: lms daemon down stops llmster. It does not stop the desktop app, and it only works when llmster is running (docs). Keep the two apart when you troubleshoot.

Step 3: Download a model and load it#

lms get downloads a model into the LM Studio model folder. The docs use openai/gpt-oss-20b as the example. They do not list sizes, so read the size in the output before you rely on a large model.

Linux#

lms get openai/gpt-oss-20b
lms load openai/gpt-oss-20b --ttl 3600 --identifier my-model
lms ps

Windows (PowerShell)#

lms get openai/gpt-oss-20b
lms load openai/gpt-oss-20b --ttl 3600 --identifier my-model
lms ps

Mac (Terminal)#

Same commands as Linux:

lms get openai/gpt-oss-20b
lms load openai/gpt-oss-20b --ttl 3600 --identifier my-model
lms ps

What the flags do, from the docs:

• --ttl 3600 unloads the model after 3600 seconds without a request. Without it, a model loaded with lms load stays in memory until you unload it.
• --identifier sets the name your apps use; the docs say it keeps the model identifier consistent.
• --gpu (max, auto, or 0.0 to 1.0) sets how much work goes to the GPU; --context-length sets the context size at load time.

Check it worked. lms ps lists the models in memory. The docs do not show its output, so look for your model in the list.

Tip: Set the context length when you load the model, not in every request. A report (issue 2533, open October 2026) describes a request with context_length that reloaded the model on every call.

Step 4: Start the server on port 1234 and check it#

The server turns the loaded model into an API. Pin the port so your scripts always know where to connect.

Linux#

lms server start --port 1234
lms server status

Windows (PowerShell)#

lms server start --port 1234
lms server status

Mac (Terminal)#

Same commands as Linux:

lms server start --port 1234
lms server status

Check it worked. The docs' example output:

The server is running on port 1234.

Now ask the API which models it serves. On Linux and Mac:

curl http://localhost:1234/v1/models

On Windows PowerShell use curl.exe: the curl name there is an alias for a different PowerShell command.

curl.exe http://localhost:1234/v1/models

The docs show this command but not its output. You should see a JSON list that includes your model. If it is empty or the connection is refused, see Troubleshooting.

Tip: Without --port, the server uses the last port you used, not always 1234 (docs). After a port change, check with lms server status before you update your scripts.

Step 5: Send a first chat request#

The OpenAI-compatible chat endpoint is /v1/chat/completions. The model field takes the identifier you set, or a model id from /v1/models.

Linux#

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "my-model", "messages": [{"role": "user", "content": "Say hello in five words."}]}'

Windows (PowerShell)#

$body = @{
  model    = "my-model"
  messages = @(@{ role = "user"; content = "Say hello in five words." })
} | ConvertTo-Json -Depth 5
Invoke-RestMethod -Uri http://localhost:1234/v1/chat/completions -Method Post -ContentType "application/json" -Body $body

Mac (Terminal)#

Same commands as Linux:

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "my-model", "messages": [{"role": "user", "content": "Say hello in five words."}]}'

Check it worked. You get a JSON reply with a choices list holding the model's message. The docs publish no sample reply; the shape is the standard OpenAI chat format.

For the official OpenAI Python library, the docs set the base URL to http://localhost:1234/v1. Their example also passes the key lm-studio; the docs do not say whether that value is checked when authentication is off.

Step 6: Decide who can reach it#

By default the server listens on 127.0.0.1: only this computer connects. To let other devices use it, bind it to all interfaces. The docs warn that any bind other than 127.0.0.1 "exposes the server beyond localhost; we recommend enabling authentication." Turn on authentication (Step 7) first.

Linux#

lms server stop
lms server start --bind 0.0.0.0 --port 1234
lms server status

Windows (PowerShell)#

lms server stop
lms server start --bind 0.0.0.0 --port 1234
lms server status

Windows Firewall may ask about the connection; allow it only on private networks you own.

Mac (Terminal)#

Same commands as Linux:

lms server stop
lms server start --bind 0.0.0.0 --port 1234
lms server status

Check it worked. From a second device on the same network, replace <server-ip> with this computer's LAN address:

curl http://<server-ip>:1234/v1/models

If you did not turn on authentication, go back to the default with lms server stop, then lms server start --port 1234. The environment variable LMS_SERVER_HOST also sets the bind address (docs).

Step 7: Require an API token#

Authentication is off by default. The switch and the tokens live in the LM Studio app: Developer tab, then Server Settings. The docs describe no command-line way to create tokens.

Linux#

The docs only describe creating tokens in the app's interface, so this step needs the desktop app on some computer. Then:

1. Open the Developer tab, then Server Settings.
2. Turn on Require Authentication.
3. Click Manage Tokens, then Create Token, and give it a name and permissions.
4. Copy the token now; the docs say it is not shown again.

Test it:

export LM_API_TOKEN="<your-token>"
curl http://localhost:1234/v1/models -H "Authorization: Bearer $LM_API_TOKEN"

Windows (PowerShell)#

The same four steps in the app, then:

$env:LM_API_TOKEN = "<your-token>"
curl.exe http://localhost:1234/v1/models -H "Authorization: Bearer $env:LM_API_TOKEN"

Mac (Terminal)#

The same four steps in the app, then the same commands as Linux:

export LM_API_TOKEN="<your-token>"
curl http://localhost:1234/v1/models -H "Authorization: Bearer $LM_API_TOKEN"

Check it worked. With the header the models list comes back; without it the request is refused. The docs do not show the error text or code, so do not rely on a specific code.

Tip: Give each app its own token with only the permissions it needs; the docs let you edit permissions per token. Keep tokens in an environment variable or a password manager, never in a shared script.

Step 8: Start it when the computer boots#

Linux#

The docs' systemd unit, saved as /etc/systemd/system/lmstudio.service. Replace YOUR_USERNAME:

[Unit]
Description=LM Studio Server

[Service]
Type=oneshot
RemainAfterExit=yes
User=YOUR_USERNAME
Environment="HOME=/home/YOUR_USERNAME"
ExecStartPre=/home/YOUR_USERNAME/.lmstudio/bin/lms daemon up
ExecStartPre=/home/YOUR_USERNAME/.lmstudio/bin/lms load openai/gpt-oss-20b --yes
ExecStart=/home/YOUR_USERNAME/.lmstudio/bin/lms server start
ExecStop=/home/YOUR_USERNAME/.lmstudio/bin/lms daemon down

[Install]
WantedBy=multi-user.target

Then enable it:

sudo systemctl daemon-reload
sudo systemctl enable lmstudio.service
sudo systemctl start lmstudio.service
systemctl status lmstudio

The docs note you can leave the model out of ExecStartPre and rely on just-in-time loading.

Check it worked. systemctl status lmstudio shows the unit active, and curl http://localhost:1234/v1/models answers after a reboot.

Windows (PowerShell)#

Not documented for llmster on Windows. The documented route is the desktop app: enable the setting that runs the LLM server on login. The docs say closing the window then minimises the app to the system tray while the server keeps running.

Mac (Terminal)#

Not documented for llmster on Mac (the docs have no launchd example). Use the same desktop-app setting as on Windows.

Check it worked. After a restart, run lms server status and expect the port line.

Security and testing#

• Keep it local by default. 127.0.0.1 is the safe bind. Change it only on purpose.
• Authenticate before you bind. Turn on Require Authentication and create a token per app (Step 7). The settings also include a switch that lets API clients call the servers you defined for the app; the docs call it a security risk and require authentication for it.
• Do not publish it to the internet. The docs document no TLS and no proxy recipe for the LM Studio server. For access away from home, use a network and VPN you control; do not forward port 1234 on your router.
• Secrets: the only secret is the API token. If you save it in a file, make the file readable only by your user (chmod 600 <file> on Linux and Mac; on Windows, restrict it in the file's properties).
• Functionality test: lms server status names port 1234, curl http://localhost:1234/v1/models lists your model, and one chat request answers (Steps 4 and 5).
• Logs: lms log stream shows incoming requests; --source server gives HTTP logs (docs). Logs can contain prompts, so treat them as private.

What leaves your machine#

• Chats and documents: the offline page says nothing you enter into LM Studio when chatting with a model leaves your device, and documents you add for chat stay local.
• Local server: requests from your apps stay on your machine or network while you keep the server off the internet.
• Model downloads and search: these go over the internet, to Hugging Face and LM Studio's catalog. The privacy policy says the only usage data it receives comes from model search or download and update checks.
• Updates: the app checks for updates, and lms daemon update downloads new versions.
• Optional cloud features: Bionic cloud models and web search are separate features with their own data terms (pricing and privacy pages). This guide does not use them.

Common mistakes#

These come from the official warnings and from issues people opened on the LM Studio bug tracker. The last row is about AI assistants.

Tip: Run lms server status after every start or update. It is the one command that tells you the port and that the server runs.

Tip: Use lms log stream when a request does not do what you expect. The docs call it useful for debugging prompts sent to the model.

Tip: Keep JIT loading on with a TTL, and let the idle timer unload models you have stopped using. The docs set JIT on by default and the idle TTL to 60 minutes.

Troubleshooting#

Exact error text#

Rehydrated error - SyntaxError: Unexpected token: the app fails to start on 0.4.26 on Windows (issue 2483). Update to the current release.

Waking up LM Studio service...: the docs' example output for lms server start. Normal in the first seconds.

Known bugs#

Open on the LM Studio bug tracker when we read it in October 2026. The pages we read give no fix; the workarounds come from the reports.

• Issue 2500 (Windows, Bionic 1.1.8): LAN clients lost access after the server switched to its own configuration. Workaround: after an update, re-check the bind from another device.
• Issue 2485 (macOS, Bionic 1.1.8): inference stuck on "changing ownership"; lms ps timed out when LM Studio was not installed. The reporter says a rollback to 1.1.7 fixed it.
• Issue 2533 (per-request context_length): reloads on each request. Workaround: set the context at load time.
• Issue 2530 (0.4.26, Windows): the API rejected a native plugin even with pluginUse: allowAll.

Most of these reports concern the desktop app and its Bionic variant, but they touch the same server.

Updates and backups#

Update. Stop llmster, update it, and start it again (the docs' steps, the same on all three systems):

lms daemon down
lms daemon update
lms daemon up

Use lms daemon update --beta only for the beta channel. Afterwards run lms server status, make a test chat request, and re-check the bind if you use the network.

Back up the models. The docs point to the models folder you set in the app's My Models tab but give no backup procedure. These commands are ours: they archive a folder you choose, with the date in the name. Replace <models-folder>.

mkdir -p ~/backups
tar -czf ~/backups/lmstudio-models-$(date +%F).tar.gz -C "<models-folder>" .

New-Item -ItemType Directory -Force "$HOME\backups" | Out-Null
Compress-Archive -Path "<models-folder>\*" -DestinationPath "$HOME\backups\lmstudio-models-$(Get-Date -Format yyyy-MM-dd).zip"

Mac (Terminal): same as Linux. Schedule the Linux or Mac command with cron at 03:00 (crontab -e), and the Windows command with a daily Task Scheduler task.

Restore test. Unpack one archive into an empty folder, point LM Studio at it, and run lms ls. The model should be listed. A partly copied model behaves like issue 2538, so compare file sizes.

Performance and resources#

The docs publish no llmster benchmark, and we measured nothing for this page. The controls the docs describe:

• GPU share and context: lms load --gpu and --context-length. A larger context uses more memory.
• Estimate first: lms load --estimate-only <model> prints the estimated GPU and total memory (API changelog, 0.3.27).
• Idle unload: JIT-loaded models unload after 60 minutes idle by default; Auto-Evict keeps one JIT-loaded model in memory at a time.
• Monitoring: lms ps and lms ps --json show loaded models and their status; lms log stream --source model --stats adds tokens per second (API changelog, 0.3.26).

The docs recommend 16 GB RAM and at least 4 GB dedicated VRAM on Windows, and publish no figures for memory per model.

FAQ#

Does it need a GPU? No. The docs describe GPU offload as an option (--gpu), and recommend at least 4 GB of dedicated VRAM on Windows. CPU-only use is not described as a supported mode on the pages we read, so test with a small model first.

Is LM Studio open source? The app is not. Its terms give a non-exclusive licence for personal and internal business use, forbid modifying, reverse engineering and redistributing it, and forbid offering it as a service. The lms command-line tool is MIT licensed.

Is it free? The pricing page lists a Free plan that includes local model support, the Bionic agent and LM Link for up to 5 devices. Paid plans exist for cloud features. Read the current pricing page before you rely on a paid feature.

Can I use it for work? The app terms allow personal and internal business use, and forbid offering the software to others as a hosted service. Each model has its own licence, and the terms say third-party models come under their own provider agreement. Check the licence of every model you serve.

Why does the app appear under two names? The pricing page, the lms README and many bug reports say "Bionic", while the docs say "LM Studio". We followed the docs and did not confirm a rename. Check the name on the download page before you install.

Is there a Docker image? Not documented. The official pages we read describe none, so this guide does not give one.

Sources#

• LM Studio docs: lms CLI - read October 2026
• LM Studio docs: run as a service (headless) - read October 2026
• LM Studio docs: llmster as a Linux startup task - read October 2026
• LM Studio docs: local API server, server settings and serve on local network - read October 2026
• LM Studio docs: authentication and idle TTL and auto-evict - read October 2026
• LM Studio docs: OpenAI compatibility - read October 2026
• LM Studio docs: system requirements, offline operation and LM Studio, llmster and lms - read October 2026
• LM Studio API changelog and release 0.4.27 - read October 2026
• LM Studio app terms, privacy policy, pricing and brand kit - read October 2026
• lms command-line tool on GitHub (MIT licence and README) - read October 2026
• LM Studio bug tracker: issues 2483, 2485, 2500, 2514, 2515, 2530, 2533, 2538 - read October 2026

Checked against the official docs on 2026-10-30.