Model Selection
Contop uses a multi-model architecture with three distinct model roles. You can configure each independently.
Three Model Roles
| Role | Where It Runs | Default | Purpose |
|---|---|---|---|
| Conversation Model | Mobile (phone) | gemini-2.5-flash | Classifies user intent, generates conversational responses, polishes execution results |
| Execution Model | Server (desktop) | gemini-2.5-flash | Controls the autonomous ADK agent for multi-step task execution |
| Computer Use Backend | Server (desktop) | omniparser | Screen understanding and UI element detection |
Available Providers
The execution model supports multiple LLM providers via LiteLLM routing:
| Provider | Example Models | API Key Required |
|---|---|---|
| Google Gemini | gemini-2.5-flash, gemini-2.5-pro, gemini-3.1-pro-preview, gemini-3-flash-preview, gemini-2.5-flash-lite | gemini_api_key |
| OpenAI | openai/gpt-5.4, openai/gpt-4.1, openai/gpt-4.1-mini, openai/o3, openai/o4-mini | openai_api_key |
| Anthropic | anthropic/claude-opus-4-6, anthropic/claude-sonnet-4-6, anthropic/claude-haiku-4-5 | anthropic_api_key |
| OpenRouter | Grok, Devstral, Qwen, Nemotron, Phi-4, MiniMax, and any model on OpenRouter | openrouter_api_key |
Non-Gemini models must use the provider prefix (e.g., openai/gpt-5.4, anthropic/claude-sonnet-4-6). Gemini models use bare names. Community models from providers like Groq, Mistral, and DeepSeek are accessed through OpenRouter.
Switching Models at Runtime
Change models from the mobile app's AI Settings without restarting:
- Open the session menu → AI Settings
- Select your preferred model for each role
- Changes take effect on the next command
Thinking Mode
Toggle extended thinking for supported models (Gemini 2.5/3.x, OpenAI o3/o4-mini, Claude Opus/Sonnet):
- Enabled - Model uses chain-of-thought reasoning (slower but more accurate for complex tasks)
- Disabled - Faster responses for simple tasks
- Default - Uses the model's built-in default
Cost/Capability Tradeoffs
| Model Tier | Speed | Cost | Best For |
|---|---|---|---|
| Flash (e.g., Gemini Flash) | Fast | Low | Simple tasks, quick commands |
| Pro (e.g., Gemini Pro, GPT-5.4) | Medium | Medium | Complex multi-step tasks |
| Large (e.g., Claude Opus) | Slow | High | Nuanced reasoning, code generation |
Subscription Mode
Instead of API keys, you can use your existing LLM subscription (Claude Pro/Max, Gemini Pro, ChatGPT Pro). In subscription mode, requests route through a local CLI proxy on the desktop that wraps the provider's official CLI tool.
- The mobile app shows a SUB badge on model chips when subscription mode is active for that provider
- A NO KEY badge appears when no API key is configured for a provider (the model can still be used via subscription)
- The mobile
use_subscriptionflag is authoritative - the phone decides per-request whether to use subscription or API key mode - Vision limitation: CLI tools accept text only - in subscription mode, the execution agent's LLM vision fallback (direct screenshot analysis) is unavailable. The agent relies on local vision backends instead.
See Configuration - Subscription Mode for setup instructions.
Vision Backends
The computer use backend determines how the agent understands your screen:
| Backend | Speed | Accuracy | Notes |
|---|---|---|---|
| OmniParser | Medium | High | Default; local-first, privacy-preserving, GPU/CPU adaptive |
| UI-TARS | Fast | High | Single API call via OpenRouter |
| Gemini Computer Use | Medium | High | Gemini-native computer use with stateful history |
| Accessibility Tree | Fast | Variable | Deterministic, best for native apps |
| Kimi / Qwen / Phi / Molmo / Holotron | Fast | Variable | Alternative VLMs via OpenRouter |
Related: Configuration · Agent Execution · ADK Agent