pi-olla-autodetect

pi-olla-autodetect

Pi extension that auto-discovers models from an Olla gateway and registers them as a provider.

Why

In a homelab with multiple inference hosts, models start and stop on demand, endpoints change constantly, and maintaining a static models.json with dozens of direct provider entries becomes a chore.

pi-olla-autodetect removes that maintenance burden. Instead of wiring each backend directly into Pi, you deploy one Olla gateway in front of your inference nodes and let this extension discover whatever models are currently available.

┌──────────────┐
│  Pi agent    │  ← sees one stable "olla" provider
└──────┬───────┘
       │
       ▼
┌──────────────┐
│  Olla gateway │  ← handles health checks, failover, routing
└──────┬───────┘
       │
  ┌────┼────┐
  ▼    ▼    ▼
sparky sparkz twins

Benefits

• One stable endpoint — Pi points to Olla, not to individual inference hosts
• Automatic model discovery — models appear in Pi as soon as Olla reports them; no models.json edits
• No stale entries — when a model goes away, it disappears from Pi on /reload
• Health checks handled by Olla — offline backends are dropped without client-facing errors
• Works alongside Open WebUI — both Pi and Open WebUI can share the same Olla gateway

Install

pi install npm:pi-olla-autodetect

For local development:

pi install /path/to/pi-olla-autodetect

Then run /reload.

How It Works

At startup (and on /reload), the extension fetches {baseUrl}/models from your Olla instance and registers all returned models under an olla provider via pi.registerProvider(). No manual model entries needed.

Configuration

Setting	Env var	Default	Description
baseUrl	OLLA_BASE_URL	http://127.0.0.1:40114	Olla server URL (any form accepted — host, host+prefix, host+prefix/v1 — normalized to /v1)
providerName	OLLA_PROVIDER_NAME	olla	Provider name registered in pi
apiKey	OLLA_API_KEY	no-api-key-needed	API key sent to Olla

Settings file: ~/.pi/olla/settings.json

{ "baseUrl": "http://127.0.0.1:40114/olla/openai/v1" }

Or override via environment variables:

export OLLA_BASE_URL=http://pluto:40114/olla/openai/v1

The extension normalizes any input format — bare host, host with prefix, or full /v1 URL — so all of these work:

OLLA_BASE_URL=http://pluto:40114
OLLA_BASE_URL=http://pluto:40114/olla/openai
OLLA_BASE_URL=http://pluto:40114/olla/openai/v1

Update Model List

Run /reload in Pi to re-fetch the model catalogue from Olla.

Requirements

• pi >= 0.74.0
• A reachable Olla gateway with model_registry.enable_unifier: true

License

MIT