The MeghaOS assistant is provider-agnostic. You pick the AI that powers it in Settings → AI, and every request routes to your chosen provider. Switching is live: the next request uses the new provider/model with no restart.

​
Any OpenAI-compatible API

The important part: most providers speak the OpenAI-compatible chat-completions format, so MeghaOS works with almost any model service — including ones not in the built-in list. There’s a dedicated Custom OpenAI-compatible API option where you supply the endpoint URL, key, and model, so you can point MeghaOS at any compatible gateway or self-hosted server. A few providers use their own native formats (Anthropic, Google Vertex AI, Amazon Bedrock); MeghaOS handles those directly.

​
Built-in providers

Choose any of these in Settings — the preset fills in the right endpoint for you.
Run models entirely on your own machine — nothing leaves the device:
  • Ollama
  • LM Studio
  • vLLM
  • Google Vertex AI — uses a service account
  • Amazon Bedrock
Custom OpenAI-compatible API — point MeghaOS at any endpoint that implements the OpenAI chat-completions API. Supply the URL, key (if needed), and model name.

​
How routing works

All calls are non-blocking, so MeghaOS stays responsive while the model thinks.

​
Live configuration

AI settings apply on the next request — no restart. Change provider, key, or model in Settings → AI, or via POST /api/settings for scripting:

​
Vision (multimodal)

Providers that support image input power MeghaOS’s vision features — screen analysis and camera/screenshot understanding. MeghaOS sends images in the right format for your provider (Anthropic vs. OpenAI-compatible).

​
Where the AI is used

Every intelligent step uses your configured model:

UI composition

Generates the A2UI interface description (plus optional JIT code).

Workflow planning

Decomposes requests into a step graph.

Memory extraction

Pulls durable facts from conversations in the background.

Direct access

POST /api/llm exposes your configured model for your own scripts.

​
Choosing a model

For the richest generated UIs and reliable tool planning, use a strong instruction-following model. Capable models from any of the providers above work well; local providers (Ollama, LM Studio, vLLM) keep everything on-device with no API key and no data leaving your machine.