Send your first chat

With a model pulled and assigned to a slot, you can chat through the prewired web UI or directly against the OpenAI-compatible API.

With OpenWebUI

hal0 installs and prewires OpenWebUI, which binds port 3001 and points at the local hal0 API out of the box.

1. Make sure the OpenWebUI unit is running:

   systemctl status hal0-openwebui

2. Open http://localhost:3001 (or http://hal0.local:3001) in your browser.

3. Pick your model from the model selector and send a message.

Note

The dashboard discovers the Chat link via GET /api/config/urls, which returns the OpenWebUI URL and whether the unit is active. On a stock LAN install it advertises http://<host>:3001. Behind a reverse proxy, set HAL0_OPENWEBUI_PUBLIC_URL in /etc/hal0/api.env — OpenWebUI emits root-absolute asset URLs and must be served on its own (sub)domain, not a path prefix.

With curl

The API is OpenAI-compatible at http://localhost:8080/v1. Use a slot name (e.g. agent) as the model id — hal0 rewrites the slot alias to the slot’s configured model before dispatching, so you don’t need to know the underlying model id.

The Endpoints tab — local OpenAI-compatible slot endpoints and slot-as-model aliases.

Non-streaming

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agent",
    "messages": [
      {"role": "user", "content": "Say hello in one sentence."}
    ]
  }'

Streaming

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agent",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Say hello in one sentence."}
    ]
  }'

The response is a Server-Sent Events stream of OpenAI-shaped chunks.

Slot aliases

hal0’s two canonical seeded roles are agent (the capable default, and the fallback anchor for every other role) and utility (a cheaper model for lightweight/background work). Both are valid model ids, as is the name of any other slot you’ve created yourself. The legacy alias agent-hermes still resolves to agent; chat and primary were retired as slot names/aliases. Passing the underlying model id directly works too, and so do the live-resolving virtual names hal0/agent, hal0/utility, and hal0/npu.

No Authorization header is required — hal0 ships with no built-in auth and treats the local network as trusted (front it with a reverse proxy if it is exposed). You can point any OpenAI SDK at the same base URL:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-needed")
resp = client.chat.completions.create(
    model="agent",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

Beyond chat

The same /v1 surface exposes embeddings, rerankings (also reachable at rerank, for clients that expect llama-server’s / Jina/Cohere-style native path), audio/transcriptions, audio/speech, and images/generations — each routed to the appropriate slot. GET /v1/models lists everything the configured slots and upstreams advertise.