LibreChat

LibreChat is an open-source, self-hostable ChatGPT-style chat UI. Docker-deployed, YAML-configured, multi-tenant, with endpoint switching, conversation history, custom prompts, and plugin support. It speaks OpenAI, Anthropic, Google, and arbitrary OpenAI-compatible providers through its custom endpoint type — the door Melious walks through. If you want a chat UI for a team without running your own LLM gateway, LibreChat plus Melious is a short recipe: one endpoint block in librechat.yaml, one .env key, one docker compose up.

Setup

MELIOUS_API_KEY=sk-mel-<YOUR_API_KEY>

version: 1.3.5
cache: true

endpoints:
  custom:
    - name: "Melious"
      apiKey: "${MELIOUS_API_KEY}"
      baseURL: "https://api.melious.ai/v1"
      models:                                                         
        default:                                                      
          - "glm-5.1"
        fetch: true
      titleConvo: true
      titleModel: "current_model"
      modelDisplayLabel: "Melious"
      dropParams: ["user"]                                            
      # iconURL: any reachable HTTPS SVG; omit for the default LibreChat icon 

services:
  api:
    volumes:
      - ./librechat.yaml:/app/librechat.yaml

docker compose down
docker compose up -d

Picking models for the default surface

LibreChat exposes a lot of models to end users. For a shared instance, curate the default list rather than dumping the whole catalog. glm-5.1 is a safe single default for general chat. Add specialized picks (code, reasoning, long-context, small/fast) from melious.ai/hub/models as your team's needs surface.

Let fetch: true keep the full list available for power users, but name the default list explicitly.

Multi-endpoint patterns

If you want Melious alongside a self-hosted Ollama or another cloud provider:

endpoints:
  custom:
    - name: "Melious"
      apiKey: "${MELIOUS_API_KEY}"
      baseURL: "https://api.melious.ai/v1"
      models: { default: ["glm-5.1"], fetch: true }

    - name: "Ollama (local)"
      apiKey: "ollama"
      baseURL: "http://ollama:11434/v1"
      models: { default: ["llama3.2"], fetch: true }

Users switch between them in the endpoint selector. We stay the default for anything cloud-hosted; Ollama handles local work.

What's different

• No prompt caching cross-user. LibreChat's cache setting is for metadata/YAML reload; per-prompt caching (as OpenAI and Anthropic expose) isn't routed through us.
• Vision inputs work for vision-capable models. LibreChat passes image blocks unchanged.
• Agents are first-class. LibreChat ships an Agent Builder with execute_code, file_search, actions, web_search, artifacts, ocr, and context capabilities. Tool calls flow through our OpenAI-compatible tool_calls schema unchanged. Pick models with strong tool support — check _meta.capabilities.tool_use on GET /v1/models?include_meta=true.

When it breaks

• Endpoint missing from dropdown — librechat.yaml wasn't mounted. Check docker compose config to see the effective volumes.
• fetch: true returns empty — your API key doesn't have the inference.models scope. Add it in the dashboard or drop fetch: true and manage default manually.
• Models show but selection fails with 404 — model IDs in default are stale. Let fetch: true do the work, or rebuild the list from GET /v1/models.

Errors and retry patterns: Errors.