Use LibreChat with open-source models hosted in the EU

LibreChat is an open-source, self-hosted chat application. Beyond conversations with many model providers, it offers agents with tools and file search, MCP server support, a code interpreter, conversation branching, presets, and enterprise sign-in through OAuth, OpenID Connect or LDAP. Organisations use it as a central assistant that they control.

Providers are configured in librechat.yaml. The endpoints.custom list accepts any API that speaks the OpenAI format, and entries with provider: anthropic speak the Anthropic Messages format. inference.eu works with either, so it appears to users as one more endpoint in the model menu.

Each custom endpoint needs a name, an apiKey and a baseURL. Reference the key as ${INFERENCE_EU_API_KEY} so it stays in .env, and set baseURL to the inference.eu base URL ending in /v1. Under models, default lists the models users always see, and fetch: true loads the full list from the API.

titleConvo and titleModel control automatic conversation titles; use a fast model so titles do not slow chats. modelDisplayLabel sets the name shown next to answers. With Docker, librechat.yaml must be mounted into the container, usually through docker-compose.override.yml, or LibreChat will ignore it.

For general chat, drafting and summarising, gpt-oss-120b is a fast default and a good title model. For long documents, file search and questions that need careful reasoning, offer glm-5.3 and kimi-k3. Teams that work with code benefit from qwen3-coder-480b-a35b-instruct.

LibreChat agents call tools, search files and use MCP servers, so assign them a model with dependable tool calling, such as kimi-k3 or glm-5.3. Model specs let admins publish curated presets, for example a contract review assistant on glm-5.3, so users do not have to choose models themselves.

If the endpoint does not appear, LibreChat did not load librechat.yaml: check the mount and the logs for a validation error. A 401 means INFERENCE_EU_API_KEY is missing from .env or the container was not restarted after adding it. A 404 usually means baseURL is missing /v1.

If the model list is empty, check fetch: true or list models under default. If a request fails with a 400 error because of an unsupported parameter, add that parameter to dropParams. If agents do not call tools, choose a model with stronger tool calling for the agent.

With inference.eu as the endpoint, every message, file excerpt and tool result sent to the model is processed in EU data centres and never used for training. We sign a data processing agreement under Article 28 GDPR, publish our subprocessor list and bill in euros.

LibreChat stores conversations, files and users in its own MongoDB database and file storage, so host them in the EU. Features such as file search embeddings, web search, speech and image generation use separately configured providers; choose EU-hosted options for them, and remove endpoints users should not reach.