Local API Server reference

The Local API Server exposes an OpenAI-compatible HTTP API on 127.0.0.1, so most existing OpenAI client libraries work by simply pointing them at your local base URL and key. This is the reference for its endpoints, parameters, and limits. To turn it on and get your key, see Set up the Local API Server.

Base URL & authentication

• Base URL: http://127.0.0.1:8756/v1 (confirm the port in Settings → Local API Server).
• Auth: every request needs Authorization: Bearer YOUR_API_KEY.

Endpoints

Chat completions

Supported request fields:

curl

curl http://127.0.0.1:8756/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Summarize: ..."}],
    "temperature": 0.7,
    "stream": false
  }'

Python (official OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8756/v1",
    api_key="YOUR_API_KEY",
)

resp = client.chat.completions.create(
    model="",  # empty = use the model loaded in the app
    messages=[{"role": "user", "content": "Summarize: ..."}],
)
print(resp.choices[0].message.content)

JavaScript (official OpenAI SDK)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://127.0.0.1:8756/v1",
  apiKey: "YOUR_API_KEY",
});

const resp = await client.chat.completions.create({
  model: "",
  messages: [{ role: "user", content: "Summarize: ..." }],
});
console.log(resp.choices[0].message.content);

Behavior & limits

• One request at a time. There’s a single on-device engine, so requests are queued. Each waits up to ~60 seconds for its turn; if it can’t start in time you get a 503 (with a Retry-After hint).
• Queue cap. When too many requests are already waiting, new ones get 429 (“server busy”). Design your client to retry with backoff — this is a personal/automation endpoint, not a high-throughput service.
• Model loading. If nothing is loaded and you don’t name a model, you get a 503 asking you to load one. Naming an installed model triggers a switch before the request runs.
• Errors are returned in OpenAI’s error shape (invalid_request_error / rate_limit_error / service_unavailable) with standard HTTP status codes (400, 429, 503).

What it’s not

• Not remote-accessible. It binds to 127.0.0.1 by design — it can’t serve other machines on your network, and that’s intentional for privacy.
• Not a multi-user server. The single-engine queue means it’s built for your own scripts and tools, not for serving many concurrent users.