Model Management

Tiered Models

Each agent in anna has three model fields, stored in the database (settings_agents table). The format for all model fields is provider/model (e.g. anthropic/claude-sonnet-4-6).

provider is usually the provider instance ID. As a compatibility fallback, anna also accepts a provider type alias such as anthropic or openai when there is exactly one configured provider of that type.

Both model_strong and model_fast fall back to model when not set. Configure these per-agent through the admin panel (anna --open).

Provider Setup

Providers are configured through the admin panel (anna --open). Each provider is stored as a plugin in the settings_plugins table (kind=provider) with credentials in the config JSON field.

Environment variables serve as fallbacks when a provider's api_key field is empty in the database:

The OpenAI-Compatible provider (openai-response) supports any service that implements the OpenAI Responses API, such as Perplexity or Together.ai.

CLI Commands

anna models             # List available models (alias for list)
anna models list        # List all models grouped by provider
anna models update      # Fetch models from provider APIs and update cache
anna models current     # Show active provider/model
anna models set <p/m>   # Switch default agent's model (e.g. anna models set openai/gpt-4o)
anna models search <q>  # Search models by name

Model Cache

anna models update queries all configured provider APIs and saves results to the settings table under the models_cache key. The cache is used by list, search, and the Telegram model picker.

You can also refresh the cache from the admin panel.

Runtime Switching

Models can be switched at runtime without restarting:

• CLI: /model command during a chat session
• Telegram: Inline keyboard model picker
• CLI command: anna models set provider/model updates the default agent's model in the database

Model Metadata

When the model cache is populated (via anna models update or the admin panel), each model entry includes metadata fetched from the provider API:

• Model ID
• Reasoning capability
• Supported input types (text, image)
• Context window size
• Max output tokens
• Cost per token (input, output, cache read, cache write)
• Custom headers

This metadata is used for model resolution, display, and cost tracking.