AI backend

The AI backend is Everlock's built-in operator assistant and shared vision runtime.

It is exposed in two places:

• an interactive prompt mode inside the admin SSH session
• image captioning for the image backend

Inference runs on Everlock's own gguf-runner engine, embedded right in the build: the model ships inside the Everlock binary, the runtime stays in-process and on CPU, and the backends share one loaded model instead of starting their own workers. Prompts and answers stay on your own machine.

What it is for

Use the AI backend when you want Everlock itself to help with operator work:

• ask questions from inside the admin shell without leaving the session
• let the model call explicit admin tools instead of free-typing commands
• generate captions and metadata for image libraries without a second AI stack

Its role is operational assistance, not general-purpose hosted chat.

How it fits in the system

The SSH path is transport plus service working together:

• frontend-ssh handles authentication, channels, and session framing
• backend-admin-ssh provides the admin shell
• backend-ai-ssh adds the prompt mode
• everlock-ai-runtime owns the shared model and worker thread, wrapping the embedded gguf-runner engine

That same runtime is reused by backend-image-http for captioning.

Every build embeds MiniCPM-V 4.6 Thinking, including the container image. See the embedded model for what it is and why its vision half is the larger one, or the engine page for how it is embedded.

When to use it

• turn it on when operators want a conversational layer over the existing admin command surface
• keep it off if you want a strictly command-only admin shell
• use it with the image backend when you want local caption generation for stored media

Pages

• The embedded model
• The gguf-runner inference engine
• AI runtime and access model
• Image backend

Read next

• Admin backends
• Image backend
• Architecture