Custom OpenAI-compatible Server

Connect MLJAR Studio to your own local or remote model server using an OpenAI-compatible API.

Before you start

Activate a lifetime MLJAR Studio license to configure your own AI provider. Open AI Provider Settings in MLJAR Studio to choose a provider.

Start your model server using its own installation guide. You need its API base URL, a chat-capable model, and an API key if authentication is enabled. For built-in presets, follow the LM Studio, Jan, vLLM, or llama.cpp guide.

Required API support

• Model discovery and connection testing require GET /models relative to the base URL, returning an OpenAI-compatible data array with model id values.
• The default Chat Completions format uses POST /chat/completions. Select Responses only if the server implements POST /responses.
• The server must support the request and streaming response formats used by your workflow. Agent tasks also depend on model and server support for tool calling.

Configure the connection

1. Open AI Provider Settings and select Custom OpenAI-compatible.
2. Set Base URL to your server API root. The preset starts with http://127.0.0.1:8000/v1; replace it with your actual endpoint.
3. Enter the server key in Custom OpenAI-compatible API key (optional) if required.
4. Choose Chat Completions or Responses in OpenAI API format to match the server.
5. Click Refresh models if necessary, then select a model from the list.
6. Click Test connection and then Save provider. Confirm the saved provider in the sidebar.

Use the base URL only: for example, an API rooted at /v1 should end in /v1, not /v1/chat/completions or /v1/models. Studio appends the endpoint path.

Test a notebook request

After saving, send a short request in your notebook. Test connection checks access to the model list and the selected model; a successful test does not guarantee that every model supports the chat or tool features your workflow needs.

Local and remote servers

OpenAI-compatible describes the API format, not where the model runs. A loopback URL reaches the machine running the notebook backend; a remote URL sends prompts and context to that server. Use HTTPS for remote endpoints. A server key comes from your server operator; an OpenAI account is not needed for a local server.

Troubleshooting

• An empty model list can mean /models is unavailable, the API root is incorrect, authentication failed, or no model is served.
• For 404 or unsupported endpoint errors, check the base URL and API format.
• If Test connection passes but generation fails, check server logs, chat model support, streaming, context limits, and any required tool-calling configuration.

Related guides

All LLM providers · Local vs Cloud LLMs · LLM Setup Troubleshooting