Run one command and your machine serves inference to the people you choose — your own laptop today, your team's grid tomorrow, the open network next. Every session your machine serves earns Vector Points. Encashment is coming; your balance starts now.
Your machine's serving is measured the same way everything in VectorGrid is — from the grid's own telemetry, visible to you in your console. No estimates, no trust-us dashboards.
Workers serve whoever you point at them — your LAN, your team. Sessions are AEAD-encrypted; stop serving with Ctrl-C. Nothing runs unless you start it.
Vector Points accrue from day one. When encashment launches, your history counts — early producers aren't starting from zero.
Any machine with spare compute — a gaming PC, an old workstation, a Mac on a desk.
$ curl -sSL https://raw.githubusercontent.com/VectorGrid/vectorgrid/main/install.sh | sh
$ vectorgrid loginSign-in links the machine to your account — that's where the points go.
$ vectorgrid analyseExact math on which models fit this machine's GPU and RAM — and what it can contribute to a pool of N machines.
$ vectorgrid worker \
--listen vgt://0.0.0.0:19310 \
--announce--announce makes drivers on your network find it automatically. Your console's Producer tab starts counting.
Two numbers matter: days your machine serves, and sessions it carries. Here's what steady serving looks like — sessions depend entirely on how busy your grid is, so we show the part we can promise and the part we can't.
| Serving pattern | Uptime points / month | + per session served |
|---|---|---|
| Weekends only (~8 days) | 800 pts | +10 each |
| Evenings (~20 days) | 2,000 pts | +10 each |
| Always on (30 days) | 3,000 pts | +10 each |
Formula v1 (preview): 100 pts per active worker day + 10 pts per secure session served. Computed server-side from the grid's telemetry — the same number your console shows. It may evolve before encashment; accrued balances carry forward.
$ vectorgrid doctorChecks GPU init, the models directory, LAN discovery, firewall rules (with the exact allow commands for your OS), WiFi power-save, and cloud-VM broadcast quirks. --fix applies the safe ones.
Workers listen on UDP 19310. On Ubuntu with ufw: sudo ufw allow 19310/udp. Announcing uses LAN broadcast — some networks (and all cloud VPCs) don't forward it; use --nodes vgt://host:19310 on the driver instead.
$ vectorgrid doctor --worker vgt://192.168.1.20:19310Probes the worker end-to-end and classifies exactly what's wrong — unreachable, firewalled, security mismatch, or healthy.
If another app pins your VRAM (games, other model servers), the worker refuses sessions honestly rather than swapping to disk. vectorgrid analyse shows what fits with what's currently free.
Only the VectorGrid worker you started, serving layers of the model a driver asks for. It executes model math — never arbitrary code from peers. Sessions are AEAD-encrypted, and the worker serves whoever can reach the port you opened: on a LAN, that's your network.
Soon — we're building the accounting and anti-abuse review that real value requires, and we won't launch it half-done. Points accrue from today and balances carry forward. No dates promised; watch your console.
No. Uptime points are per day a machine actually served, and session points require a real driver connecting. Before encashment, balances go through an anti-abuse review — gaming the preview only wastes your electricity.
A worker only computes when a session is active, and it refuses sessions that don't fit free memory. You can stop any time — drivers fail over automatically.
Whoever you give network reach: your own machines, your team's LAN. The open peer-to-peer network — serving strangers, with matching rewards — is the roadmap Vector Points exist for.