ln-mcp

MCP server that gives AI agents a Bitcoin Lightning wallet. Pay invoices, fetch L402 APIs, discover paid services, and manage spending budgets — all through the Model Context Protocol.

Works with Claude Desktop, Claude Code, Gemini CLI, and ChatGPT.

Demo project: Aegis Wallet — a seedless Bitcoin wallet where Claude is the AI financial agent. Uses ln-mcp for autonomous Lightning payments with budget enforcement and real-time approval flows.

Quick Start

npx -y ln-mcp

Claude Desktop

Add to Claude Desktop → Settings → Developer → Edit Config:

{
  "mcpServers": {
    "ln-mcp": {
      "command": "npx",
      "args": ["-y", "ln-mcp"],
      "env": {
        "LND_MACAROON_BASE64": "<your scoped macaroon>",
        "LND_REST_HOST": "https://localhost:8080",
        "NODE_TLS_REJECT_UNAUTHORIZED": "0"
      }
    }
  }
}

Claude Code

claude mcp add ln-mcp \
  -e 'LND_MACAROON_BASE64=<macaroon>' \
  -e LND_REST_HOST=https://localhost:8080 \
  -e NODE_TLS_REJECT_UNAUTHORIZED=0 \
  -- npx -y ln-mcp

Gemini CLI

gemini mcp add \
  -e 'LND_MACAROON_BASE64=<macaroon>' \
  -e LND_REST_HOST=https://localhost:8080 \
  -e NODE_TLS_REJECT_UNAUTHORIZED=0 \
  ln-mcp npx -- -y ln-mcp

ChatGPT (HTTP mode)

HTTP_PORT=3004 LND_MACAROON_BASE64=<macaroon> LND_REST_HOST=https://localhost:8080 npx -y ln-mcp
# Then tunnel: ngrok http 3004
# Paste the ngrok URL + /mcp into ChatGPT Settings → Apps → Create

Tools (10)

How It Works

AI Agent (Claude / Gemini / ChatGPT)
    ↕ MCP (stdio or HTTP)
ln-mcp
    ↕ HTTPS + macaroon
LND REST API (your node)
    ↕ gRPC
402 Index API (service discovery)

The MCP server connects directly to your LND node's REST API. The macaroon controls what the agent can do — scoped permissions + budget ceiling enforced cryptographically by LND.

Payment Flow

User: "Buy me a coffee"

Agent: search_services(q="coffee", protocol="L402")
  → Finds Unhuman Coffee (L402, ~3000 sats)

Agent: l402_discover("https://unhuman.coffee/api/order")
  → Gets pricing, parameters, usage docs (free)

Agent: l402_fetch("https://unhuman.coffee/api/order", method="POST", body=...)
  → Hits URL → gets HTTP 402 + invoice
  → Decodes invoice (3000 sats)
  → Pays via Lightning (SendPaymentV2, 60s timeout)
  → Caches L402 token for unhuman.coffee
  → Retries with L402 auth header
  → Returns order confirmation + receipt

Budget Exceeded → User Approval

When a payment exceeds the agent's budget, the invoice is forwarded to the wallet app for the user to pay directly:

Agent: l402_fetch(url)
  → LND rejects (FAILURE_REASON_INSUFFICIENT_BALANCE)
  → MCP POSTs to AEGIS_WEBHOOK_URL: { bolt11, amount, error }
  → Wallet app receives SSE push → shows "Pay directly" banner instantly
  → User taps to approve → payment completes
  → Agent relays TELL_USER message to user

Only budget-exceeded failures trigger this flow. Routing failures and timeouts return an error for the agent to relay — no webhook, no false "approval required" messages.

Environment Variables

Transports

Requirements

• Node.js 22+
• LND or litd node with REST API enabled (port 8080)
• A macaroon with payment permissions

Recommended Macaroon Permissions

routerrpc.Router/SendPaymentV2
routerrpc.Router/TrackPaymentV2
lnrpc.Lightning/SendPaymentSync
lnrpc.Lightning/DecodePayReq
lnrpc.Lightning/ChannelBalance
lnrpc.Lightning/ListPayments
lnrpc.Lightning/GetInfo
lnrpc.Lightning/AddInvoice

For budget enforcement, use a litd account macaroon — the budget ceiling is enforced cryptographically by LND's RPC middleware.

Design Decisions

• No get_balance tool — agents used it to pre-check balance and refuse to pay, preventing the budget-exceeded → webhook → user-approval flow. Balance is available via get_spending_summary and in every payment receipt.
• No decode_invoice tool — agents decoded invoices, saw the amount, and hesitated before paying. pay_invoice and l402_fetch decode internally.
• SendPaymentV2 streaming — 60-second timeout auto-cancels stuck payments. No more "already in flight" errors on retry.
• Webhook only on budget exceeded — routing/timeout failures return an error to the agent directly, not a false "manual approval required" message.

License

MIT