Skip to content

FAQ

No. All ML in Alcoves is CPU-only by design — face/object detection, audio tagging, and speech transcription all run on CPU. Models are selected for that constraint. See Privacy & local AI.

No. There is no telemetry, no analytics, and no inference that leaves your instance. The Go backend is a pure API that never phones home.

A machine with roughly 8 GB of RAM is the practical ceiling for the CPU-only models. Storage scales with your library — use a local disk or any S3-compatible object store. A Raspberry Pi with a USB drive and a rack with an object store are both supported.

With Docker and Docker Compose for a single host, or the Helm chart for Kubernetes. External PostgreSQL and Dragonfly/Redis are expected; storage is local or S3. Start with the Quickstart.

The first registered user becomes the owner. Owner-gated surfaces include the admin panel, the job-queue dashboard, registration policy, and ML-model selection. Within libraries, the roles are owner / admin / viewer.

Alcoves is alpha (0.x.y). Expect breaking changes — they ship with migrations and changelog entries. It targets bounded, trusted instances (1–100 users), not anonymous mass signup.

Yes — Alcoves ships an MCP (Model Context Protocol) server so MCP-capable clients can browse libraries and move files over stdio or HTTP, authenticated by a personal access token.