Cloud Providers

Cloud providers give Mobile Agent access to the largest and most capable models. You connect a provider once — with OAuth or an API key — and its models appear in the composer picker for any conversation.

Provider families

Transports

Each model uses a transport that decides how the app talks to the provider:

• openaiChat — OpenAI Chat Completions (/chat/completions).
• openaiResponses / codexResponses — OpenAI Responses API (including Codex-aware endpoints).
• anthropic — Anthropic Messages API.
• google — Google Generative Language API.
• openaiCompatible — your own base URL.
• onDevice — the LiteRT-LM local engine.

Connect a provider

1. Open Settings → Providers.
2. Tap the provider you want.
3. OAuth providers (Google, OpenAI, etc.) start a secure sign-in flow and store the resulting account — no token handling on your side. API-key providers (Anthropic, xAI, OpenRouter) just need a key pasted into the field. Ollama needs only the server address.
4. Mobile Agent discovers available models and lets you set a default preset plus per-conversation overrides.

When you're done, the provider shows a connected status. A failed connection surfaces the error so you can fix it (bad key, network, or blocked endpoint).

Custom OpenAI-compatible profiles

For any gateway, proxy, or self-hosted stack, use OpenAI-compatible and set the base URL. The app lets you choose the transport:

https://your-gateway.example.com/v1

Both Chat Completions and Responses are supported. This is the standard way to reach vLLM, LM Studio, Together, Groq, local tunneled servers, or corporate endpoints.

Per-conversation model selection

Models are selected per conversation, not globally:

• Use the composer's model picker to switch the current conversation to another provider/model on the fly.
• Compare models inline — context window, tool support, image input/generation, and reasoning capability are shown for each.
• The ModelPreset configuration stores which models each provider exposes plus your default.

This makes it easy to use a fast on-device model for quick chats and a frontier cloud model for hard problem-solving in the same session.

Multi-modal support

When the active model supports it, the composer accepts image attachments (image input) and the agent can generate images — for example with an image-capable OpenAI model. The resolved model capabilities are checked automatically before offering these options; a model that can't see or draw won't pretend it can.

Reasoning effort

Models with reasoning support expose a ReasoningEffort from none → xhigh (none, minimal, low, medium, high, xhigh). Select the level per conversation to trade depth against latency and cost. Structured thinking, especially for planning and debugging, benefits from high/xhigh.

Privacy model

Mobile Agent itself never sees or proxies your traffic. Applies to cloud providers specifically:

1. You type a message in the app.
2. The app sends that conversation's messages and tool definitions directly to the provider you chose.
3. The provider streams tokens back.
4. Only usage summaries (token counts, cost) are stored on-device per message.

Turn off all cloud providers and stick to on-device models for a 100% offline setup — see On-device Models.