FYRHostVPS for AI Agents — Always-On, Token Signup, x402
Always-on boxes for autonomous agents: bearer-token account, x402 USDC or 19-coin top-up, open API, no KYC. Scout from $3/mo funded.
An agent holds a wallet and speaks HTTP; the hosting it needs is a persistent KVM it can fund and administer programmatically. The path: mint a token over the API, top it up (x402 USDC on Base or any of nineteen coins), deploy from balance — root on a plan sized to the fleet, no KYC anywhere in the loop.
What an agent actually needs from a host
Three things, and none of them is a dashboard. A credential the agent can hold — a bearer token minted over plain HTTP, standing in for the inbox-plus-password a human types. A balance it can fund itself — a signed stablecoin transfer or a catalog coin, credited by a watcher, no facilitator. And a box that stays up — a persistent KVM whose state, disk and IP survive between the agent's API calls, which hourly microVM markets do not sell. The x402 page documents the rail end to end: x402 VPS carries the machine-flow specifics.
VPS from $3* (fund $50+), storage 1–5 TB NVMe from $24/mo annual-eff, dedicated with IPMI from $30.80/mo. Crypto invoice, no KYC.
Launch nowPlaybook: mint, fund, deploy
- POST /api/agent/account — the token is the account; minting accepts the AUP.
- POST /api/agent/topup with X-Payment-Mode: x402 — sign EIP-3009, broadcast, retry with the tx hash; or fund with BTC/XMR/USDT like a human.
- GET /api/agent/catalog — plans, locations, premiums, priced exactly as the human checkout shows.
- Deploy from balance: plan, flag, OS; root credentials return as JSON.
Sizing an agent fleet
The same ladder humans use, read by machines. A monitor or a Telegram-grade bot fits Scout ($5.49 list, $3/mo funded): one vCPU, 2 GB ECC, enough for a heartbeat loop. Sentry ($7.50 / $6.00) is the production floor — 2 vCPU / 4 GB carries a working agent with local state. Fleets consolidate onto Ranger or Warden; anything queueing disk writes belongs higher or on the dedicated line from $30.80/mo annual-eff. Model weights up to ~30 GB are a storage-tier workload, not an inference ticket — this catalog sells no GPUs, and the honest section says what that means.
Where agents actually run
Observed fleets on this pattern: uptime and change monitors that need always-on egress, trading and arbitrage loops close to exchange endpoints, Telegram/Matrix bot hosts, scrapers with PTR discipline, CI runners spawned by repo webhooks, and community-funded nostr relays. The common shape is small, persistent and stateful — which is exactly the slice-and-ECC shape of the catalog, not the container-by-the-hour shape of the x402 sandbox market.
Security model for machines that own root
An agent with a token has full control of its box and its balance, so the blast radius of a prompt injection or a poisoned dependency is the whole account. Operational hygiene that actually helps: least-privilege API keys in the workload, egress allowlists where the fleet's job permits them, the balance topped up to the fleet's weekly need rather than a quarter's, and alerts on balance draw-down anomalies — a compromised agent spends money in patterns its owner did not write. The AUP binds the mint: whatever the agent does, the account owner accepted.
Limits
No GPU inference here — routing, gateways, bot logic and consensus-adjacent tooling are the honest workloads, and model hosting means weights on disk, not training runs. x402 settles USDC on Base specifically; other assets take the coin path. Payments are final — no facilitator means no reversal. And the AUP is payer-agnostic: CSAM, phishing, spam and paid attacks are banned no matter what mints the token. Machine-grade honesty lives on the x402 page; this page is the workload view of the same rails.
Does an agent need an email address?
No. The account is a bearer token minted over the API; an inbox is optional and only matters for human-style deployment mail.
Can one token run a whole fleet?
Yes — one balance funds many deployments; the catalog endpoint prices each before the agent commits.
What stops a rogue agent from spending everything?
Nothing but the operator's design: top-up to weekly need, alerts on draw-down, and revocation by stopping the top-ups — there is no chargeback on any rail.
Do agents get a different AUP?
No. One policy covers both checkouts: CSAM, phishing, spam and paid attacks are banned no matter who — or what — pays.
Can a human take over an agent's box?
Yes — the token works in the same panel humans use; SSH keys and credentials are the agent's to share with its owner.
Primary sources
Ready to launch?
Build the box — VPS, storage or bare metal — create the password, pay the invoice that follows.