In ShadowCode: a bundled runtime.
ShadowCode ships a pinned llama.cpp runtime under /usr/lib/shadowcode/ that runs GGUF models on Vulkan GPUs or the CPU. Settings › Local models shows the runtime, the detected hardware, your models and a short list of free Apache-2.0 models to download on request, with one recommended for your memory and graphics card. Nothing downloads until you choose Download, and each download is checked against a pinned SHA-256. Models in an existing Ollama store can be imported by reference without copying. You do not need Ollama or LM Studio.
No model weights ship on the ISO. With no subscription and no API key, ShadowCode's first run offers a free model that fits your computer, and until you connect a subscription the model picker lists the free models on this computer first. Without a Vulkan GPU the runtime falls back to the CPU; a model that runs out of graphics memory offers a smaller context or another model. The runtime listens on 127.0.0.1 with a new random key for each launch.
In Mission Control: the on-device provider.
Mission Control chooses a provider per mission. localmodel carries the code_change and sourced_report capabilities (its bridge is chat-only, so in 5.0.0 it writes cited reports but does not change code) and runs with network policy none. Its sandbox gets its own network namespace, so it reaches neither the internet nor the host's loopback services; it reaches the model over a unix domain socket that Firebreak bind-mounts into the sandbox read-only.
That socket is a deliberate opening: the model service on the other side runs outside the sandbox, as your desktop user. Until you supply such a service, the installation reports its local state as installed-unavailable, and available only when a registered on-device provider actually is. Where more than one provider is ready for a mission's capability, pass --provider.
Keep working offline.
Project workspaces, local files, creative applications and deterministic media exports need no AI provider. shadowfetch-agent-network offline starts agent sandboxes with no network; ShadowCode has its own Offline mode that runs only models on this computer.
Measurements.
This site publishes no performance figures for 5.0. The earlier model measurements are historical records of one card and one runner, not claims about ShadowCode's runtime.