What it needs

The floor is any machine with 2 GB of RAM and 2 CPU cores, amd64 or arm64, with an external AI provider: a mini PC, a NAS, a small VM or a cheap VPS. Kept’s images are built for both architectures, so Docker on Apple Silicon and ARM cloud servers run them too.

A Raspberry Pi with 2 GB or more will probably work, but it is not a promise: Kept isn’t tuned for it, and no release is tested on one.

Not yet measured on the floor itself

Kept’s performance targets are promised on this floor: 10,000 things in a location (50,000 in an instance), search under 300 ms and a thing’s page under 200 ms of server time at the 95th percentile, under 400 MB of memory for the web app and worker together, and ready within 10 s of a cold start. They have been measured on a laptop limited to the floor’s size; the run on a real 2 GB, 2-core VM, once on amd64 and once on arm64, is still due.

What needs more

• Local AI with Ollama. A model running on the same machine shares its memory with Kept, and it hasn’t been measured on the floor yet. Until it is, treat Ollama as something for a bigger machine; see Ollama.
• Photos. Kept resizes uploaded photos as they arrive. With under 3 GB of RAM it resizes one at a time, else two (KEPT_IMAGE_CONCURRENCY overrides it).
• Disk. The database is small; files are what grow. Leave room for backups on another disk, and Admin → Status warns, and admins are alerted, when a disk passes 85 % full.

Postgres

Kept needs PostgreSQL 18 with pgvector. Compose runs the pgvector/pgvector image, pinned by digest; for a provider’s database see managed Postgres.