Requirements and sizing

For operators who self-host. Reference.

Software

Item	Needed for	Version or note
Python	Backend, worker, command line	3.13 (requires-python is >=3.13,<3.14)
uv	Installing the backend with uv sync --frozen	Any recent release
Node and npm	The web app (npm ci, npm run dev, npm run build)	A recent Node. The web image builds on Node 22.
Docker with Compose v2	The compose stack	docker compose
NVIDIA container toolkit	The GPU Decider profile	Needed on the host for docker-compose.gpu.yml
strands-decider	A Decider server outside Docker	Version 0.1.0, installed with pip install strands-decider==0.1.0

Replay mode needs only Python, uv and Node. It makes no network call to OpenStreetMap, the Decider or Amazon Bedrock, and needs no GPU or AWS account.

Hardware

Setup	What it runs on
Replay mode	Any machine that runs Python and Node.
Decider on GPU	An NVIDIA GPU. The AWS template uses a g6.xlarge (4 vCPU) and also allows g5.xlarge.
Decider on CPU	A CPU only host. The AWS template allows a c7i.2xlarge (8 vCPU). The CPU compose file estimates 2 to 30 seconds per request, so use it for background analysis only.

The README says CPU latency has not been measured.

Memory minimums have not been measured. No test in the repository records the RAM the CPU Decider uses or the GPU memory the 2B model needs, so this page gives no minimum. What is known:

• The model weights are about 5 GB on disk, and the Decider loads them into memory when it starts. Plan for at least that much free memory, plus room for the API and the operating system.
• The AWS GPU host, a g6.xlarge, has one NVIDIA L4 with 24 GB of GPU memory and 16 GiB of host memory. The CPU host, a c7i.2xlarge, has 16 GiB of memory. The Decider has run on the GPU host; the CPU host has not been timed.
• The AWS template raises an alarm when GPU memory use passes 90 percent, because the Decider may no longer answer at that point. Watch that figure on your own hardware.

If you measure a smaller setup, please report it in an issue.

Disk

What	Size
Decider model weights, downloaded on first run	About 5 GB (the README says about 4.5 GB)
Root volume on the AWS host	100 GiB by default. The base image snapshot is 75 GiB.
Database	One SQLite file. Its size grows with the number of cases.

The compose stack keeps the database in the cso-data volume and the model cache in the hf-cache volume.

Ports

Port	Used by	Reachable from
443	Caddy, the only published port on the AWS host	The internet
8080	Caddy on a local machine, with docker-compose.local.yml, mapped to port 80 inside the stack	127.0.0.1 only
8000	The API (cso api binds 127.0.0.1 by default)	Loopback only
8099	The Decider server	Loopback only. It has no ports: entry in compose.
5173	The web dev server (npm run dev)	Your machine
22	Not used. There is no SSH and no key pair.	None

The Decider server and the API listen on 127.0.0.1 only. Do not publish port 8000 or 8099.

Outbound network

To	Why
api.openstreetmap.org	Reading changesets. GET only.
maproulette.org	Importing tasks. GET only.
www.openstreetmap.org	One POST to the token endpoint, for sign-in
Amazon Bedrock in your region	Drafts, in live mode with the gate on
tiles.openfreemap.org	The browser loads map tiles for the optional case map

Next: Install with Docker Compose.