Architecture

MoleculeDesk is built as a layered engine, so the CLI and future desktop app can share the exact same logic:

CLI / Desktop
      |
    Core
      |
   Runtime
      |
Model Registry
      |
Boltz / AlphaFold / ColabFold / Chai / ...

• CLI / Desktop — moldesk parses commands and delegates to the core engine. It has no business logic of its own, which is what lets the future desktop app reuse the exact same engine.
• Core — the shared engine. It composes the runtime, adapters, and registry; owns atomic installation transactions and installed-model state; and exposes doctor, list, install, and (in progress) run.
• Runtime — everything that touches your machine: hardware detection, Python/Docker discovery, subprocess execution, isolated uv environments, and checksum-verified asset caching.
• Model Registry — validates and loads model manifests, so every model is described the same way regardless of what it does upstream.
• Adapters — thin, model-specific translations for installation and execution. Each supported model (Boltz, ProteinMPNN, LigandMPNN, ...) has one.

Why this matters as a user

You don't need to know which layer does what to use MoleculeDesk — but it explains two things you'll notice:

1. Every model is installed and listed the same way, no matter how different its upstream setup normally is.
2. A model's registry status (available, beta, planned) reflects how much of this pipeline has been verified end-to-end for that model — see Model compatibility & registry.