Brain is a set of containerised services with a message bus, durable storage
and model endpoints behind them. The surprising part is how little hardware
it needs: the models are the demanding component, and only one of them wants
a GPU.

## Hardware

Guidance by what you have, not by tier. Which inference model runs where is on
[model options](/brain/docs/on-premise/model-options).

| Your hardware | Embeddings | Reranking | Ingest models |
| --- | --- | --- | --- |
| Laptop or small server, no GPU | local (CPU) | cloud | cloud |
| Integrated GPU (Vega, Iris, Apple Silicon) | local | cloud | cloud |
| Workstation with a discrete GPU | local | local | local or cloud |

Running every inference model locally wants roughly 16 GB or more of VRAM to
hold the ingest models concurrently. Without a discrete GPU, reranking is the one
that has to move to a cloud endpoint — it is the one part of the system that
genuinely needs the hardware. Everything else runs on commodity kit.

## Services

- **A NATS server with JetStream.** One process, one command — and not
  optional: it is where the knowledge lives.
- **An embeddings endpoint.** Ollama is the cheapest local path; a cloud API
  works if you would rather not run one.
- **Durable storage.** The knowledge base is the asset; it needs to survive a
  host failure like any other system of record.
- **Somewhere to terminate TLS** and route traffic to the MCP endpoint.

## What sizes a deployment

- **Corpus size and growth** — how many documents you hold, and how many
  arrive per week. Ingest is model-bound, so this drives model capacity
  before it drives disk.
- **Concurrent agents** — how many clients query and capture at once, and
  whether they are interactive or batch.
- **Search volume** — query embedding is cheap and flat; reranking is the
  part that scales with it.
- **Network posture** — connected, restricted egress, or genuinely
  air-gapped. Air-gapped changes updates and support, not just firewall
  rules.
- **Who operates it** — your team, ours, or shared. This decides what access
  we need and where the support boundary sits.

## What does not change

The MCP endpoint, the tools and the client setup are identical to the hosted
product — the [install guides](/brain/docs/install/claude-code) work
against your own URL. Agents neither know nor care where Brain runs.

## Next

[Model options](/brain/docs/on-premise/model-options) covers the four
inference models and where each can run. For figures against your own numbers,
register below with a rough scale — that is faster than any table.