Models & keys
AgentDesk is multi-model and bring-your-own-key. The app is free; you pay your AI provider directly for usage. Add one or more provider keys and AgentDesk routes work to the right model.
Supported providers
| Provider | Default model | Key |
|---|---|---|
| Claude (Anthropic) | claude-sonnet-4-6 | ANTHROPIC_API_KEY |
| OpenAI | gpt-4.1 | OPENAI_API_KEY |
| DeepSeek | deepseek-chat | DEEPSEEK_API_KEY |
| Gemini (Google) | gemini-2.0-flash | GEMINI_API_KEY |
| Grok (xAI) | grok-3 | XAI_API_KEY |
Lightweight internal tasks — session extraction, classification, commit analysis — use a small, cheap model (Haiku) to keep costs down, while heavier reasoning uses a stronger model.
Configuring keys
Add keys in the Settings panel. AgentDesk resolves them in priority order:
- Your keys — stored encrypted (Fernet) per user.
- Server keys — set by an admin for a metered/shared deployment.
- Environment —
.envfallback for local development.
This means you can run fully bring-your-own-key (you pay each provider), or an admin can supply metered keys for a team.
Semantic memory search needs an embeddings provider (e.g. OpenAI). Add that key to get the most out of search_memory; without it, recall falls back to keyword matching.
Backend
The desktop app talks to a backend (hosted at api.biskilled.com, or your own deployment). Memory is stored in PostgreSQL + pgvector when configured; otherwise AgentDesk falls back to local file-based capture. You can self-host the backend and database for full control.
That's the tour. Start with Install, or book a demo for a walkthrough on your own codebase.