An Agent Skill that lets an AI agent provision a Solana VPS on its own — paid for in USDC over x402, no human in the loop.
If you give your agent this skill, it can run a single command and get back a freshly-provisioned Linux VM with the agent runtime of its choice (OpenClaw, Hermes Agent, or plain Ubuntu) already configured with the model and API key the agent picked. Billing is per‑VM, paid in USDC via x402; the agent's wallet is the customer account.
Cloud providers were built for humans clicking dashboards. Agents need an HTTP endpoint that takes a stablecoin and returns a server. SVMClaw is that endpoint; this repo is the skill file that makes it discoverable to OpenClaw, Hermes, Claude Code, and any other runtime that reads SKILL.md.
When loaded, your agent gains one capability:
"Provision a new SVMClaw VM (Solana VPS) paid for in USDC via x402."
The skill walks the agent through:
- picking a tier (Nano $10 / Micro $20 / Standard $80 per month),
- picking an agent runtime + model + API key (or a plain Ubuntu box),
- calling
POST https://api-claw.svmacc.tech/api/vm/agent/create, - receiving an HTTP 402 challenge with the exact USDC amount and recipient,
- paying it on Solana mainnet,
- retrying the request with an
X-PAYMENTheader, - receiving the VM's ID, IP, and SSH key.
Everything after step 1 is automatable. The agent's wallet identifies it as a customer — no signup, no API key issuance.
git clone https://github.com/svmacc/svmclaw-skill.git
bash svmclaw-skill/install.shThe installer finds your ~/.openclaw (or ~/.moltbot / ~/.clawdbot) directory, copies the skill into workspace/skills/svmclaw/, and registers it in TOOLS.md. Re-run after upgrades — it's idempotent.
git clone https://github.com/svmacc/svmclaw-skill.git ~/.hermes/skills/svmclawHermes picks up skills from ~/.hermes/skills/<name>/SKILL.md on next session.
mkdir -p ~/.claude/skills
git clone https://github.com/svmacc/svmclaw-skill.git ~/.claude/skills/svmclawClaude Code auto-loads any folder under ~/.claude/skills/ that contains a SKILL.md.
The skill is just skill/SKILL.md plus a folder of helper scripts. Drop it wherever your runtime looks for skill packs and point your agent at SKILL.md.
To actually provision a VM, the agent needs:
| A Solana wallet | The signer's public key becomes the SVMClaw customer ID. Any wallet works (Phantom, Solflare, a raw keypair file). |
| USDC on mainnet | Enough to cover one month of the chosen tier (Nano = 10 USDC, Micro = 20, Standard = 80). USDC mint: EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v. |
| The ability to sign and submit a Solana tx | Most agent frameworks can do this with a few lines of @solana/web3.js or any Solana SDK. |
You do not need an SVMClaw account, dashboard login, or pre-issued API key — the payment is the authentication.
{
"message": "Payment accepted. VM created. Provisioning started.",
"vm": {
"id": "svm-abcd12",
"status": "creating",
"tier": "nano",
"agent_image": "hermes",
"ip": null,
"billing_end": "2026-06-28T..."
},
"tx_signature": "5xyz...",
"customer_id": "cust_...",
"paid_usd": 10
}The VM's IP appears within a minute or two as provisioning finishes; the agent can poll GET /api/vm/list/:customerId (with a JWT issued via SIWS) or simply re-fetch the VM record.
Pulled live from the SVMClaw API, never hardcoded in the skill:
| Tier | Monthly | vCPU | RAM | Disk |
|---|---|---|---|---|
nano |
$10 USDC | 2 | 2 GB | 40 GB |
micro |
$20 USDC | 2 | 4 GB | 70 GB |
standard |
$80 USDC | 4 | 8 GB | 180 GB |
No add-ons. No hidden fees. The price you see in the 402 challenge is the price you pay.
The endpoint provisions one of three OS/runtime layers:
none— plain Ubuntu 22.04 or 24.04 (just a server)openclaw— Ubuntu + OpenClaw runtime, pre-configured with your modelhermes— Ubuntu + Hermes Agent (Nous Research), pre-configured with your model
Model providers supported by both OpenClaw and Hermes:
anthropic · openai · google · xai · mistral · groq · huggingface · openrouter · openai-codex (OAuth, no API key needed)
Model IDs are typically <provider>/<model> (e.g. anthropic/claude-sonnet-4-6); the skill includes a curated default list. Bare model IDs are also accepted when the provider field disambiguates routing.
- Your API key is sent to SVMClaw at provision time and injected into the VM's
~/.hermes/.envor~/.openclaw/openclaw.json. The VM then makes inference calls directly to the provider using that key. SVMClaw does not log or proxy your key, but it does briefly hold it in memory during provisioning. If that's not acceptable, useprovider: openai-codex(OAuth — you link a ChatGPT subscription on the VM, no key leaves your control). - The agent pays before the VM exists. Idempotency is handled by the on-chain transaction signature — if your retry uses the same signed tx, SVMClaw returns the same VM rather than provisioning a second one. But if your agent signs two different txs for the same logical request, that's two VMs and two charges.
- The treasury wallet is
8pYBUtSEv1vfyN1uvW3ndRaK5gsMJ1bKJR8HpQdrr3rE. Always read it from the live 402 challenge, never hardcode it — it could change. - Mainnet only. There is no devnet endpoint right now. Don't point your agent at this with mainnet USDC unless you mean it.
| ✅ Today | Provision VMs with x402 + tier/runtime/model selection. Skill works in OpenClaw, Hermes, and Claude Code. |
| 🟡 Soon | list-vms, stop-vm, start-vm, renew-vm capabilities (same endpoint family, separate scripts). |
| 🟡 Soon | Dedicated MCP server (@svmacc/mcp-server) so any MCP client (Cursor, ChatGPT, Perplexity) can use SVMClaw without a SKILL.md loader. |
| 🟡 Later | Hot-resize tier without redeploying. Tier sold-out subscribe-and-wait. |
This is the first release; behavior may shift while we learn what agents actually need. We follow semver — breaking changes bump the major.
Open an issue at https://github.com/svmacc/svmclaw-skill/issues with the exact request body, the response status, and the response body. Don't paste your API key.
MIT — see LICENSE.
SVMClaw is run by @svmacc. Frontend at https://claw.svmacc.tech, API at https://api-claw.svmacc.tech, facilitator at https://facilitator.svmacc.tech.