Connect Langfuse Playground

In an existing Langfuse project, an LLM Connection lets Playground and evaluations call a model. It is separate from tracing your application’s requests. Use an account that can manage the project’s connections. The configured service key is available through that connection to permitted project users; prompts and outputs pass through Langfuse. Review project access and data retention before using private data.

1. Add the LLM connection

Open Project Settings → LLM Connections → Add new LLM API key. Give it a distinct connection name and choose OpenAI as the provider adapter.

Enter your service key privately in the API key field, without Bearer. Under Advanced Settings, set Base URL to the service Base URL above, including /v1. Leave Use Responses API off for this OpenAI-compatible Chat Completions setup.

Add the exact service model IDs you want under Custom model names, using the authenticated models endpoint as the source. Do not assume that Langfuse’s built-in vendor model list matches the models available to your key. Save the connection without replacing an existing connection of the same name.

2. Use it in Playground

Open Playground, select the saved connection and one of its custom text models, then run a short prompt. This generates a paid model request. Confirm the model and request in your service dashboard.

The connection persists in the project. Reopen the connection or model selector after editing it. If a request returns 404, check the Base URL and protocol setting; for model errors, check the exact custom model name and key permissions.

Evaluations and tracing

For LLM-as-a-Judge evaluations, select this connection explicitly and use a model that supports OpenAI-format tool calling. Evaluate a small sample before enabling automatic runs: each judge call can spend balance.

Adding an LLM Connection does not instrument application code or change how existing traces are collected. Likewise, Langfuse’s model definitions and cost estimates are separate from your actual service billing. Review the project’s prompt and trace handling before sending private evaluation data.

See the official LLM Connections guide for adapter settings and project access requirements.