Hardware requirements¶
Daedalus is delivered as SaaS — as a tenant you run no control-plane hardware. The only machines you provide are the endpoints you enroll agents on, and those requirements are tiny. The fuller specs below are for anyone self-hosting the platform.
Tenant (SaaS) — what you provide¶
| You provide | Minimum |
|---|---|
| A browser to reach the console | any modern browser (the console is mobile-first) |
| Endpoints to test | hosts you own/are authorized to test, meeting the agent row below |
That's it — inference, the executor sandbox, the control plane, and storage are all hosted.
Agent endpoint¶
The BAS agent is deliberately lightweight; it polls out and runs bounded technique chains.
| Resource | Minimum | Notes |
|---|---|---|
| CPU | 1 core | bursts only while a chain runs |
| RAM | 512 MB free | ~1 GB if using the container runtime path |
| Disk | ~250 MB | native agent; more if the container image is pulled |
| OS | Linux, macOS, or Windows 10/Server 2016+ | Windows agent is native — no container runtime needed |
| Network | outbound HTTPS | see Network — no inbound ports |
Self-hosting the platform¶
If you run your own deployment, three roles matter. They can share a host for a lab, but separating the executor is strongly recommended (it's the only tier that runs untrusted commands).
| Role | Minimum | Recommended | Notes |
|---|---|---|---|
| Control plane | 2 vCPU · 2 GB RAM · 20 GB SSD | 4 vCPU · 4 GB RAM · 40 GB SSD | Flask + SQLite; light. State lives here — back up the DB. |
| Executor / sandbox | 4 vCPU · 8 GB RAM · 40 GB SSD | 8 vCPU · 16 GB RAM · 80 GB SSD | runs a container runtime; tool images are large, hence the disk. |
| GPU inference node | GPU with 8 GB VRAM · 16 GB system RAM | 12–16 GB VRAM · 32 GB RAM | see below. CPU-only works but is far slower. |
GPU inference detail¶
The reference node runs a tool-capable 7B model fully GPU-resident on an 8 GB consumer card, using grouped-query attention and KV-cache quantization to fit a 16k context. See Local inference for the tuning that makes 8 GB enough.
- 8 GB VRAM is the practical floor for a GPU-resident 7B at a useful context.
- 12–16 GB lets you run a larger model or a longer context comfortably.
- No GPU? The platform runs on CPU inference, but generation is slow enough that it's only suitable for testing, not real engagements. A cloud model can be configured for the executive reasoning instead.
Start small, separate later
A single 8 vCPU / 16 GB host plus one 8 GB GPU is enough to run everything for a home lab. Split the executor onto its own host before pointing agents at anything you care about — that's the tier that runs untrusted input.