sparkbtcbot-proxy

A serverless proxy that lets AI agents use a Spark Bitcoin L2 wallet over HTTP, without exposing the private key.

Spark is a Bitcoin L2 with instant payments and sub-satoshi fees. This proxy wraps the Spark SDK behind authenticated REST endpoints so agents can check balances, send payments, and create invoices — while you keep the mnemonic safe on the server.

Why use this?

If you give an agent direct SDK access (sparkbtcbot-skill), the agent holds your mnemonic. That's fine for testing, but risky in production.

This proxy solves that:

• Mnemonic stays on server — agents get bearer tokens, not keys
• Spending limits — cap per-transaction and daily spend (global or per-token)
• Revocable access — cut off a compromised agent without moving funds
• Role-based auth — give agents invoice-only access if they don't need to spend
• L402 support — pay Lightning paywalls automatically and fetch protected content

Token roles

The API_AUTH_TOKEN env var is a hardcoded admin fallback — it always works even if Redis is down. Use it to bootstrap: create scoped tokens via the API, then hand those to agents.

API

All routes require Authorization: Bearer <token>.

Notes:

• POST /api/pay, POST /api/transfer, POST /api/l402, and GET /api/l402/status require admin or pay-only token
• POST /api/invoice/create and POST /api/invoice/spark require admin or invoice token
• Token management routes (/api/tokens) require an admin token
• Read-only routes (/api/balance, /api/info, etc.) work with any role
• POST /api/l402/preview works with any role (doesn't spend)

Example: create an invoice

curl -X POST https://your-deployment.vercel.app/api/invoice/create \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"amountSats": 1000, "memo": "Test invoice"}'

Returns:

{"success": true, "data": {"encodedInvoice": "lnbc10u1p..."}}

Example: L402 paywall

L402 lets agents pay for API access with Lightning. The proxy handles the full flow: detect 402, pay invoice, get preimage, retry with auth.

curl -X POST https://your-deployment.vercel.app/api/l402 \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://lightningfaucet.com/api/l402/joke"}'

Returns:

{
  "success": true,
  "data": {
    "status": 200,
    "paid": true,
    "priceSats": 21,
    "preimage": "be2ebe7c...",
    "data": {"setup": "What's a programmer's favorite hangout?", "punchline": "Foo Bar!"}
  }
}

#### Preview L402 cost

Before paying, you can check what an L402-protected resource will cost:

curl -X POST https://your-deployment.vercel.app/api/l402/preview \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://lightningfaucet.com/api/l402/joke"}'

Returns:

{
  "success": true,
  "data": {
    "requires_payment": true,
    "invoice_amount_sats": 21,
    "invoice": "lnbc210n1p...",
    "macaroon": "AgELbGlnaHRuaW5n..."
  }
}

If the resource doesn't require payment (non-402 response), you'll get requires_payment: false with the response data.

#### Handling pending L402 payments

Lightning payments via Spark are asynchronous. If the payment succeeds but the preimage isn't available within the timeout window (~7.5 seconds), the proxy returns a pending status instead of failing:

{
  "success": true,
  "data": {
    "status": "pending",
    "pendingId": "a1b2c3d4...",
    "message": "Payment sent but preimage not yet available. Poll GET /api/l402/status?id=<pendingId> to complete.",
    "priceSats": 21
  }
}

Agents must handle this case. When you receive status: "pending":

1. Wait 2-5 seconds
2. Call GET /api/l402/status?id=<pendingId>
3. If still pending, repeat steps 1-2 (up to ~30 seconds total)
4. Once complete, you'll get the full response with the protected content

# Poll for completion
curl "https://your-deployment.vercel.app/api/l402/status?id=a1b2c3d4..." \
  -H "Authorization: Bearer $TOKEN"

The pending state is stored in Redis with a 1-hour TTL. If you don't poll within that window, the pending record expires and you'll need to make a new L402 request (which will pay again).

Important for agent developers: Your agent logic should include a retry loop for L402 requests. The payment has already been sent — failing to poll means you paid but didn't get the content.

#### Token caching

L402 tokens are cached per-domain and reused automatically. When you make a request to a domain you've already paid, the proxy tries the cached token first. If it still works, you get the content without paying again. If the server returns 402 (token expired), the proxy pays for a new token and caches it.

The response includes cached: true when a cached token was used:

{
  "success": true,
  "data": {
    "status": 200,
    "paid": false,
    "cached": true,
    "data": {"setup": "...", "punchline": "..."}
  }
}

Tokens are cached for up to 24 hours (or until the server rejects them).

#### Automatic retry for empty responses

Some L402 servers return empty or null content immediately after payment (they may not have processed the preimage yet). The proxy automatically retries the final fetch up to 3 times with 200ms delays if the response looks empty. This covers cases where the protected content has null fields (e.g., {"setup": null, "punchline": null}) or is entirely empty.

Environment variables

Getting started

You'll need a Vercel account (free tier works) and an Upstash Redis database (free tier works).

Option 1 — install as a Claude Code plugin (for using the deployed proxy)

If your goal is to give an agent access to a proxy that already exists (yours or someone else's), install the skill via the Claude Code plugin system:

claude plugin marketplace add https://github.com/echennells/sparkbtcbot-proxy
claude plugin install sparkbtcbot-proxy

That registers this repo as a marketplace and installs both the API-usage skill and the deploy skill. The agent can then call the proxy with PROXY_URL and PROXY_TOKEN env vars set.

Option 2 — clone and deploy your own proxy

git clone https://github.com/echennells/sparkbtcbot-proxy.git
cd sparkbtcbot-proxy
npm install
npx vercel --prod

Set the environment variables in the Vercel dashboard, then redeploy.

For detailed step-by-step instructions (including generating a mnemonic and creating the Redis database via API), see skills/deploy/SKILL.md. That file is also part of the Claude plugin above, so an installed agent can drive deployment for you.

See also

sparkbtcbot-skill — gives an agent direct Spark SDK access. Simpler (no server), but the agent holds the mnemonic and there are no spending limits.