Credentials & runner

These two pages control how models run (accelerators / backend) and who can call them (tokens and OpenAI-compatible URLs).

Runner

AI → Runner shows:

1. Inference backend — typically llama.cpp (llama-server) with version and a Ready pill.
2. Accelerators detected on this host — CPU, Apple GPU (Metal), Apple Neural Engine, and others depending on hardware.
3. License-gated Combine as one runner for multi-accelerator dispatch.

The orange In runner card is what new loads use. Prefer Metal on Apple Silicon for local GGUF / llama.cpp. NPU may appear as present but unused until the runner supports it.

TIP

Combining CPU + GPU + NPU into one runner is a licensed feature. Without a license, LaunchPad keeps a single accelerator in the runner (shown on the lock card).

Credentials

AI → Credentials has five blocks.

OpenAI gateway

Clients talk to LaunchPad with an OpenAI-compatible base URL and an sk-elp- API key. Inference backends stay on localhost; the gateway listens so VMs and LAN clients can reach it.

This Mac / Electros (HTTP): http://127.0.0.1:11434/v1. From a VM: https://192.168.67.1:11434/v1 (hypervisor bridge IP may vary).

Example with curl after you create a key:

curl http://127.0.0.1:11434/v1/models \
  -H "Authorization: Bearer sk-elp-YOUR_KEY"

Local network access

Turn on Share on local network (HTTPS) to expose a LAN URL. Remote clients must trust the LaunchPad CA — use Trust for Electros (then restart the app) and/or Install in Keychain. Electros does not use the macOS Keychain automatically for its own TLS trust.

Hugging Face token

Optional token for gated model downloads. Paste and Save.

Cloud providers

Add provider connects OpenAI-compatible clouds (OpenRouter, OpenAI, Anthropic, …). Each entry shows masked key, base URL, and optional include filters (for example openrouter/free). Use Filters, Refresh, or Remove per row.

API keys

Create key mints an sk-elp- secret. Edit access scopes it to all models or a subset. Revoke immediately invalidates the key — copy secrets when created; you will not see the full string again.

Checklist: expose a model to a VM

1. Load a model under Running models.
2. Credentials → Create key (all models or that model only).
3. From the VM, call the From a VM HTTPS base URL with the key.
4. If TLS fails, install / trust the CA from the LAN section, then retry.

Next

• AI models — load something to serve
• Services — Open WebUI and proxies that use these URLs