Turn LM Studio into an OpenAI API server (2026)

#What you get

By the end of this guide, you’ll have an http://localhost:1234/v1 endpoint that any OpenAI SDK (Python, JavaScript, C#), LangChain, or coding tool such as Cline or Continue can use instead of the OpenAI API. The server also provides a native API at /api/v1 (stateful chat, model loading and downloading) and Anthropic-compatible endpoints. Everything stays on your machine—the model, requests, and responses do not leave your computer unless you expose it on the network yourself.

#1. Start the server

1. 01
   Open the Developer tab

   In LM Studio, the Developer tab contains the server, its logs, and its settings. The model to be served must be downloaded beforehand.
2. 02
   Enable Start server

   Toggle the Start server switch: the server starts on the port specified in Server Settings, 1234 in the documentation examples.
3. 03
   Or start from the command line

   From a terminal, the lms server start command launches the same server without opening the interface.
4. 04
   Check the model list

   Query /v1/models to confirm that the server responds and see the model identifiers to use in your requests.

lms server start

#2. Server settings, one by one

The settings are under Developer, Server Settings. They determine who can call the server and what clients can make it do. Most integration problems come from one of these switches, usually the network or CORS setting.

#3. Test with curl

A first control call is enough to validate the server. In requests, the model field must contain the model identifier as it appears in LM Studio, not a generic name such as local-model: the documentation reiterates this in its curl example.

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

curl http://localhost:1234/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "IDENTIFIANT-DU-MODELE",
    "messages": [
      {"role":"system","content":"Tu es concis."},
      {"role":"user","content":"Capitale du Portugal ?"}
    ],
    "temperature": 0.2
  }'

The response is JSON in OpenAI format: choices[0].message.content contains the text. If the model field is wrong, or on-demand loading is disabled and the model isn't loaded, the request fails: first check the identifier returned by /v1/models.

#4. Call from Python

The openai SDK is used by changing only the base URL: this is the change shown in the LM Studio documentation. The SDK requires a key; as long as authentication is disabled, LM Studio does not check it, and if you enable authentication, the key becomes your API token.

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio",  # ignorée sans authentification ; votre jeton sinon
)

resp = client.chat.completions.create(
    model="IDENTIFIANT-DU-MODELE",
    messages=[
        {"role": "system", "content": "Réponds en 1 phrase."},
        {"role": "user",   "content": "Qu'est-ce qu'un LLM ?"},
    ],
    temperature=0.3,
    stream=True,
)

for chunk in resp:
    delta = chunk.choices[0].delta.content or ""
    print(delta, end="", flush=True)

Streaming works like OpenAI's: no client-side changes are needed to display tokens live. For agents and editors, LM Studio also implements the /v1/responses endpoint, described below.

#Which API to choose: OpenAI, Anthropic, or native

LM Studio exposes three endpoint families. OpenAI-compatible endpoints cover models, responses, chat, embeddings, and completions. Anthropic-compatible endpoints accept Anthropic’s message format. Since version 0.4.0, the native /api/v1 API adds functions specific to LM Studio: stateful chat, model loading, unloading, and downloading, and per-request context configuration.

#6. Multiple models: on-demand loading and TTL

With just-in-time loading (JIT, for Just in Time), the first call to a model loads it into memory, and /v1/models lists all downloaded models, not just the ones that are loaded. Without JIT, /v1/models returns only models that are already loaded, and you must load the model before calling it. This mode is ideal when a tool such as Zed, Cline, or Continue chooses its own model.

Default TTL

A model loaded on demand is unloaded after 60 minutes without a request.

TTL per request

Add a ttl field (in seconds) to the request; 300 corresponds to 5 minutes.

TTL for lms load

Models loaded with lms load have no default TTL: use the --ttl option.

Auto-Evict

Enabled by default: only one model loaded on demand remains in memory at a time. Disable it to keep several models loaded.

#7. Expose the server on the network with authentication

For another computer on the network to call your server, enable Serve on Local Network in Server Settings, or start it with the listening address 0.0.0.0. The server will then stop listening on localhost alone: the documentation warns that any bind other than 127.0.0.1 exposes it beyond the machine and recommends enabling authentication.

lms server start --bind 0.0.0.0

curl http://192.168.1.42:1234/v1/models

Contrary to popular belief, LM Studio can authenticate requests. By default, it requires no authentication; when you enable the toggle in Server Settings, it accepts only requests carrying a valid API token, created in Manage Tokens with selected permissions. The token is displayed only when created: copy it immediately. This feature requires LM Studio 0.4.0 or later.

curl http://192.168.1.42:1234/v1/models \
  -H "Authorization: Bearer $LM_API_TOKEN"

For access from the Internet, don’t expose the port: use a VPN or a reverse proxy with TLS. The principle is identical to that of a Ollama server, detailed in the security guide. A simpler alternative for using a model from another machine is LM Link, which serves a model from a remote device as if it were loaded locally.

• Securing a local AI server: authentication, reverse proxy, exposure

#No graphical interface: llmster and automatic startup

Since version 0.4.0, the core of LM Studio has existed as a standalone daemon, llmster, designed to run without an interface on a Linux server, a GPU machine, or a local workstation. It installs with one command line, starts with lms daemon up, and the server then starts with lms server start. On a workstation with an interface, you can also check the option in the application settings that starts the server when you log in: closing the application then minimizes it to the system tray while the server continues running.

curl -fsSL https://lmstudio.ai/install.sh | bash
lms daemon up
lms server start

#9. Performance: what really matters

GPU offload

Load as many layers as possible into VRAM. A model that spills into system memory loses most of its throughput.

Context Length

Choose the context you need, not the maximum: the context cache uses VRAM and grows with the context length.

Max Concurrent Predictions

Number of requests processed simultaneously by a model; beyond that, they wait in a queue.

Unified KV Cache

Enabled by default: resources are not divided into fixed shares between requests, allowing variable request sizes.

The Flash Attention guide and the context-window guide detail the effects on memory. For substantial multi-user throughput, a dedicated server is still better suited: the vLLM guide shows the deployment.

• Flash Attention in llama.cpp
• Understanding the context window

#Limits and alternatives: what changed

Several commonly cited limitations are no longer true, and correcting them changes the choice of tool. The table compares what you still read with what the current documentation says.

Real limitations remain: LM Studio is designed for a workstation, not a cluster; continuous batching does not replace a server designed for dozens of users, such as vLLM; and application updates can change behavior, so you must pin the version on a machine that provides a service. To choose between LM Studio and its competitors, compare them before committing.

• Deploy vLLM in production
• Ollama vs. LM Studio vs. Jan vs. GPT4All
• LM Studio on Linux
• Connect Cline to a local model

• Source: LM Studio documentation, local API server
• Source: LM Studio documentation, server settings
• Source: LM Studio documentation, authentication
• Source: LM Studio documentation, OpenAI compatibility
• Source: LM Studio 0.4.0 announcement