Docs / Connections

Connections

A connection is an endpoint and its key — one server you run, or one account at a provider. Its models are listed from the endpoint and enabled one by one.

Two types

Keys are encrypted at rest and only ever shown masked.

Add an OpenAI-compatible endpoint

1. Go to Admin → Connections → New connection.
2. Enter the address, including /v1 — for a local runner usually something like http://localhost:1234/v1 — and a key if it needs one.
3. Give it a provider name (see below).
4. Press Test & refresh. Its models are discovered and cached.
5. Enable the models you want under Admin → Models.

Add an Anthropic key

1. Admin → Connections → New connection, and choose Type: Anthropic. The address fills in https://api.anthropic.com.
2. Provider name: anthropic, or whatever the models should be called by.
3. Paste the API key from the Anthropic console, save, and enable the models under Admin → Models. If credits are in use, give them a price.

An Anthropic model works everywhere a chat model does — chats, titles, compaction, image review, schedules and /v1. A reasoning effort becomes extended thinking. Every request is marked for prompt caching on the system prompt and the newest message, so a long chat is not paid for in full on every turn. Embeddings, transcription, speech and image generation stay with OpenAI-compatible connections.

Provider names and model ids

Every connection has a provider name: short, lowercase letters, digits and hyphens — llama, deepseek, anthropic — and unique on the server. Every model is then named provider/model, for example llama/qwen or deepseek/deepseek-chat. That is the id the API, the model picker and the LLeMbas CLI all use, and it can never be ambiguous.

Model settings

Each model has a page under Admin → Models. The list is searchable and scales to hundreds of models. On a model’s page you set:

• a display name, description and image; ordering and pinning; the instance’s default model;
• tools — whether a tool list may be sent at all, and which built-in tools. An endpoint without tool support rejects the whole request, so this is a real switch;
• reasoning — the effort levels the model understands and its default;
• vision, context window and output limit;
• agent execution — whether it may drive agent chats;
• which groups may use it, and its price when credits are in use;
• a per-model system prompt and default sampling parameters.

Each person can also narrow things for themselves under Settings → Models: switch off tools a model may use with them, and pick the effort a new chat starts at. Those defaults only ever narrow what the instance allows.

Data groups

Every connection belongs to a data group, which decides what part of your library its models may read. See Permissions and sharing.