Hardware
One machine that does nothing else
Agent Office runs on one dedicated machine: a small mini PC is enough when the models run in the cloud, and a workstation with a large graphics card can run the models itself. The installer measures the machine and sizes the team's work to it.
Supported platforms
| Platform | Status |
|---|---|
| Ubuntu 24.04 LTS, x86_64, on a dedicated host (bare metal or a VM) | Supported Clean install, re-run, upgrade, uninstall and reinstall verified in a WSL2 distribution and in a KVM virtual machine. No bare-metal install recorded yet. |
| Ubuntu 24.04 LTS, aarch64 (Arm) | Untested. The installer accepts it, but no install has been run. |
| Ubuntu 22.04 LTS, Debian 12 | Not supported today: their Python is older than 3.12, and the installer stops. |
| Other Linux distributions with systemd | Best effort, untested. |
| WSL2 on Windows | Not for production. It stops when its last session ends, and it exposes the machine's local ports to Windows. Used only to verify installs. |
| macOS, Windows, containers without systemd | Refused by the installer. |
- RAM
- 4 GB minimum. RAM decides how many cards run at once; the reference boxes below have 16 GB.
- CPU
- 2 threads minimum, 4 recommended.
- Disk
- 20 GB free minimum, 50 GB free on an SSD recommended.
- GPU
- Not needed, unless you want to serve models on the same machine.
- Network
- Outbound HTTPS during the install. After that: your AI providers, your channel accounts (email, social, webhooks), the search service your employees use, and the other flows the security summary lists. With every model on your own network, model calls need no internet access.
Dedicated means dedicated. The AI employees can hold real tools on the machine, and the kill switch and firewall rules act on the whole machine. Keep personal files, other services and development work off it.
Reference box classes
The tier and parallel tasks below are what this release's own tuning code computes for a headless machine that runs nothing else. They are estimates until a device is rated. Example prices were checked on 27 September 2026 and are shown only as dated examples of each class; Agent Office LLC has no arrangement with any vendor, and memory and GPU prices rose sharply through 2026.
| Class | Reference configuration | Tier | Parallel tasks | Local models | Example price |
|---|---|---|---|---|---|
| Small office mini PC | 4 to 8 cores, 16 GB RAM, 500 GB SSD, wired Ethernet | t4 | 14 | None: use a cloud provider or a separate model host | $385 to $620 new |
| Refurbished business desktop | 6-core Intel Core i5 (8th generation or newer), 16 GB, 256 to 512 GB SSD | t4 | 14 | None | $305 to $309 refurbished |
| A VM in your own cloud | 4 vCPU, 16 GiB, general purpose (not burstable), 64 GB SSD volume or more | t4 | 14 | None on these sizes | $98 to $147 a month for compute |
| Workstation with a GPU | 16 threads, 64 GB, 1 to 2 TB NVMe, one GPU with 24 GB of memory or more, a model server installed | t5 | 32 | Full | $1,300 to $1,900 for a 32 GB card alone |
Small office mini PC
Enough for a small team when its RAM is right. Look for 16 GB of RAM (the step from t3 to t4), a wired Ethernet port, socketed RAM and an M.2 slot. It will ship with Windows; you replace it with Ubuntu 24.04 LTS. Examples: Beelink EQ14 (Intel N150, 16 GB, 500 GB) at $385; Beelink SER8 (Ryzen 7 8745HS, 16 GB, 512 GB) at $619.
Refurbished business desktop
Usually the cheapest way to a 16 GB box, with firmware suited to an always-on machine. Replace any spinning disk with an SSD, install the latest firmware, and look for a refurbisher's warranty. Examples: Lenovo ThinkCentre M920q Tiny at $305; HP EliteDesk 800 G4 Mini at $309, both with a Core i5-8500T and 16 GB.
A VM in your own cloud account
You pay the cloud bill; Agent Office LLC hosts none of it. Choose a general-purpose size that is not burstable (Kanban workers run for long stretches), x86_64, Ubuntu 24.04 LTS from the cloud's own images, and an SSD volume of 64 GB or more. The dedicated-host rule applies to a VM too, and you should lock down the cloud metadata service. Today you launch the VM, connect over SSH and run the installer; ready-made images are a later idea.
| Cloud (region) | Size | vCPU / RAM | Tier | Parallel tasks | About a month |
|---|---|---|---|---|---|
| AWS (us-east-1) | m7i.xlarge | 4 / 16 GiB | t4 | 14 | $147 |
| Azure (East US) | D4s v5 | 4 / 16 GiB | t4 | 14 | $140 |
| Google Cloud (us-central1) | e2-standard-4 | 4 / 16 GB | t4 | 14 | $98 |
On-demand Linux compute only, checked 27 September 2026; disk, snapshots, data transfer and tax are extra, and commitments cost less. Sizes with 2 vCPU reach t3 (8 tasks); 8 vCPU reach t5 (32 tasks).
Local models: which GPU for which model
Agent Office needs no GPU. A local model is optional, and every call to it goes through the inference gateway like any cloud model, with the same data policy, budgets, metering and kill switch. Agent Office never ships, bundles or downloads model weights. ao local-models plan prints what the machine could hold and installs nothing; you install the model server and the weights yourself, under each model's own licence.
The sizing assumes the 65,536 tokens of context every employee-facing model class requires, which makes the cache a large share of the memory.
| GPU memory | Verdict | What fits entirely on the GPU | Realistic use |
|---|---|---|---|
| None, 8 GB or 12 GB | None or light | Nothing at 64K context | Keep models on a cloud provider or a separate host |
| 16 GB | Light | Nothing; 8B and 14B models split across GPU and system RAM | The fast or worker class for routine cards, preferably on a separate host |
| 24 GB | Full | An 8B model at 8-bit or a 14B model at 4-bit | The worker and fast classes, or a small lead |
| 32 GB | Full | A 14B model at 8-bit, a 24B model at 4-bit, or an 8B model at 16-bit | A lead-class model of about 24B, or a worker model with room to spare |
| 48 GB | Full | A 32B model at 5-bit, a 24B model at 8-bit, or a 14B model at 16-bit | Lead and worker from 24B to 32B models; room for a second model |
Example cards, checked 27 September 2026: a used NVIDIA GeForce RTX 3090 (24 GB) at about $1,450; an AMD Radeon AI PRO R9700 (32 GB) at $1,299 list and $1,400 to $1,900 in stores (check ROCm support on Ubuntu 24.04 first); an NVIDIA RTX PRO 4000 Blackwell (24 GB) at about $2,250 to $3,030.
Consider two boxes. Run Agent Office on a small box, and serve models from a separate GPU host on your network. The model host can be rebooted and upgraded without stopping the office. In this release a model server on the same box reads as attested either way, so you give up little.
Whether a model is good enough for a job is not something GPU memory decides: the gateway checks each deployment's context and tool use, and you judge the work. Token rates depend on the model and runtime you choose, so we do not publish them.
Hardware ready
A box is Hardware ready when it passes every item below on a fresh install of a named release. No device has been rated yet. The first rating planned is a refurbished business desktop in the 16 GB class, installed bare metal and headless, with its outputs published unedited.
- Architecture: x86_64.
- CPU: 4 threads or more.
- RAM: 8 GB or more (tier
t3); 16 GB recommended (t4), within the vendor's stated maximum. - Storage: an SSD, with 50 GB or more free after the install; 1 TB or more if the box will keep model weights.
- Operating system: Ubuntu 24.04 LTS from the latest point release, with every device working on Ubuntu's own kernel.
- Network: a wired Ethernet port that works with no extra driver, and a fixed or reserved address. Wi-Fi alone does not count.
- Firmware: the latest version, power restore after an outage on, AHCI storage mode, the integrated GPU's memory share at its minimum, a firmware password set.
- Always on: never suspends, and a synchronised clock (sign-in codes and audit timestamps depend on it).
- Nothing shared: headless, with no desktop session, database, container runtime or anyone else's work on it.
- Swap: 2 GB of swap, zram, or both.
- The result:
ao specsexits cleanly with every minimum met, andao doctorreports 0 failed.
Ratings will publish the tested configuration, how we got the device, a dated price with its source, and the four command outputs. No rankings, no vendor logos, and no payments that change a rating. A small UPS is recommended, not required.
Rate your own box
After installing, run:
sudo ao --company <id> specsThe Tier, Concurrency and Limited by lines are the rating. When your box gets less than its class, Limited by says why: ram means more RAM (or less of it taken by a desktop or other services), cpu means more threads, and tier means the box has reached its class.