
Nwc Mcp
io.github.llmops-pro由 GitHub 账号 llmops-pro 发布v0.2.3更新于 Oct 24, 2026
Lightning wallet for AI agents via Nostr Wallet Connect (NIP-47): budget caps, confirm, audit log.
概览
基于 Nostr Wallet Connect 的闪电钱包 MCP 服务器,让 AI 助手查询余额、创建发票并支付 sats,并带有预算上限与审计日志。
- 可以做什么
- 查看钱包与交易:能力信息、余额、按支付哈希或 bolt11 查询发票、本地解析 bolt11、预算状态
- 按指定金额生成收款发票
- 支付 bolt11 发票、Lightning Address、LNURL 字符串,以及一次最多 20 张的批量支付
- 向公钥发起 keysend 支付,仅在显式开启时可用
- 使用一次性令牌确认待执行的支付
- 每次支付调用都会经过只读、白名单、预算与二次确认等检查,并写入 JSON 审计日志
- 适用场景
- 适合让助手自主支付小额真实 sats 的场景,例如为付费工具调用付款、给动态打赏、按任务支付费用,或向客户开具收款发票,同时把支出限制在你设定的范围内。
- 运行要求
- 以本地 Node.js 20+ 进程运行(npm 包 nwc-mcp,stdio)。需要一个支持 NIP-47 的钱包(如 Alby Hub、Mutiny、Coinos、Phoenix)提供的 NWC 连接字符串,通过 NWC_CONNECTION_STRING 传入;未设置 NWC_DAILY_BUDGET_SATS 时服务器拒绝启动。可选变量用于调整上限、只读模式、二次确认、keysend、白名单与日志路径。需要访问钱包中继的网络连接。
安装
在 SourceWeft 中
- 打开 控制台中的 Nwc Mcp,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
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
Configure
Copy .env.example to .env and fill in your NWC connection string:
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:
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:
NWC_READ_ONLYgate — if set, blocks at the first gate. No decode, no network, audit log entry only.- Invoice decode + amount resolution — for bolt11-based tools. Rejects ambiguity (amount-override conflicting with embedded amount).
- Allowlist check — if
NWC_ALLOWED_DESTINATIONSis 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. - Budget check — daily rolling cap + optional total cap + optional per-invoice cap. State is persisted to
NWC_BUDGET_STATE_PATH. - Two-step confirmation gate — if
NWC_REQUIRE_CONFIRM=true, returns a 16-byte hex token with 5-minute TTL instead of executing. Agent must callnwc_confirm_paymentwith the token. Safety re-runs on consume. - Execute — calls the wallet via NIP-47. On success: budget records principal + (over-rounded) fees.
- 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:
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
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:
[email protected]— 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.
来源:README.md,提交 d405e54
工具
0版本历史
1- v0.2.3最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分