nwc-mcp

A Lightning wallet for your LLM agent. MCP server that exposes Nostr Wallet Connect (NIP-47) wallet operations as tools your agent can call — read balance, generate invoices, pay invoices, pay Lightning Addresses, pay LNURLs, batch payments, keysend — all wrapped in a defense-in-depth safety stack.

Drop it into Claude Desktop, Claude Code, Cursor, or any MCP-speaking client. Hand the agent a sub-wallet (not your main one). Set a daily cap. Let it spend sats on your behalf within rails you control.

v0.2 — all spend tools shipped. Twelve tools (7 read/receive + 5 spend + 1 confirm), four safety gates, full audit log. Validated end-to-end against Alby Hub + Olympus by ZEUS routing — real outbound payment settled in 8 seconds.

What you can do with this

• A coding agent that pays for its own paywalled tool calls (L402-style paid MCP servers).
• A NOSTR bot that zaps interesting notes under a daily cap you set.
• An evaluation harness that pays a small fee per task to a worker agent it consumes.
• A storefront agent that issues receive invoices in response to incoming DMs.
• Personal sats-aware agents — "every Friday, send 1,000 sats to my favorite OSS maintainers via their LN addresses."

The safety stack is the load-bearing reason this is usable in production: the agent can request a payment, but the server refuses anything that breaches your configured limits, demands a second-step confirmation if you set one, and writes every call to a structured audit log.

The twelve tools

Read-only (always enabled, no spend risk)

Receive (safe — only generates invoices, no spend)

Spend (gated by the safety pipeline)

Requirements

• Node 20+
• A NWC connection string from a NIP-47-compatible wallet (Alby Hub, Mutiny, Coinos, Phoenix, …)

The canonical buyer setup is Alby Hub PRO with a dedicated sub-wallet for the agent — that's what this server has been smoke-tested against. But the protocol is wallet-agnostic; anything that speaks NIP-47 works.

Install

# From npm (once published)
npx -y nwc-mcp

# From source
git clone <repo>
cd nwc-mcp
corepack enable pnpm
pnpm install
pnpm build

Configure

Copy .env.example to .env and fill in your NWC connection string:

cp .env.example .env
# edit .env: NWC_CONNECTION_STRING=nostr+walletconnect://...

The server auto-loads .env from the current working directory and from the project root (one level up from dist/). Env vars passed by the parent process (e.g., claude mcp add --env ...) take precedence over .env. .env is gitignored — keep your NWC string out of any committed file.

Required

Optional safety knobs

Wire into an MCP client

Claude Code (project-scoped, recommended)

From this directory:

claude mcp add nwc -s project node "$(pwd)/dist/index.js"

This writes a .mcp.json in the project root. Project-scoped servers load only when Claude Code starts from this directory tree. Use /mcp inside a session to inspect what's loaded. Restart the CLI to pick up a fresh dist/index.js after a rebuild.

Switch -s project to -s user if you want the server available everywhere.

Claude Desktop / Cursor / other clients

See examples/claude-desktop-config.json for the manifest snippet. Copy it into your client's config, replace the path placeholder. Because the server loads .env itself, you don't need to pass NWC_* vars through the client config — leave the env block empty.

Safety model

This server is designed for use by autonomous LLM agents that hold the keys to spend real sats. The defense in depth runs in this order on every spend tool call:

1. NWC_READ_ONLY gate — if set, blocks at the first gate. No decode, no network, audit log entry only.
2. Invoice decode + amount resolution — for bolt11-based tools. Rejects ambiguity (amount-override conflicting with embedded amount).
3. Allowlist check — if NWC_ALLOWED_DESTINATIONS is set, the payee (pubkey for bolt11, LN address for LUD-16, LNURL string for LUD-01, pubkey for keysend) must match. Empty allowlist = no-op.
4. Budget check — daily rolling cap + optional total cap + optional per-invoice cap. State is persisted to NWC_BUDGET_STATE_PATH.
5. Two-step confirmation gate — if NWC_REQUIRE_CONFIRM=true, returns a 16-byte hex token with 5-minute TTL instead of executing. Agent must call nwc_confirm_payment with the token. Safety re-runs on consume.
6. Execute — calls the wallet via NIP-47. On success: budget records principal + (over-rounded) fees.
7. Audit log — every attempt (success, blocked, error) is appended as one structured JSON line per call.

The floor is your NWC connection's own daily cap, set in your wallet. This server's checks are belt-and-suspenders on top — if you trust nothing about the server, the wallet still won't spend past its connection-level cap.

Verifying calls actually went through

Independent of what the LLM tells you, tail the audit log:

tail -n 5 nwc-mcp.log

Successful call: {"ts":"...","tool":"nwc_pay_invoice","outcome":"ok","result":{"paid_sats":50,...}}. Blocked: {"ts":"...","tool":"nwc_pay_invoice","outcome":"blocked","blocked_reason":"..."}. The audit log is append-only by intent; rotate it as part of your operational hygiene.

Testing

pnpm typecheck   # tsc --noEmit
pnpm test        # 25 vitest cases covering Allowlist, ConfirmStore, evaluateSafety, decodeLnurl
pnpm build       # produces dist/index.js (~47 KB ESM bundle)

evaluateSafety is a pure function that the unit tests exercise across the full safety matrix (read-only / amount-conflict / allowlist / budget / confirm) without touching a real wallet.

For end-to-end smoke testing against a live wallet, see paywall-mcp-test/ in the parent project — a throwaway second MCP server that exposes one paid tool (premium_compliment) and validates the full agent → invoice → pay → redeem loop.

License

MIT — see LICENSE.

Contact / Issues

Built by LLMOps.Pro.

• NOSTR: npub1hdg932jvwc3jdvkqywgqv0ue4nn60exrf92asy8mtazt3hjg7d2s2yw0nw — follow, DM, zap.
• Lightning Address: sovereigncitizens@getalby.com — for support zaps and "this was useful" tips.
• Bug reports / feature requests: open a GitHub issue (https://github.com/llmops-pro/nwc-mcp).
• Security issues: please disclose privately via NOSTR DM before opening a public issue.