Spec Driven Development
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
作者提供的说明(英文原文)
Self-custodial pump.fun sniper: arm strategies, snipe, manage positions. Simulate by default.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
SNIPER_MODE, SNIPER_NETWORK, SOLANA_RPC_URL, SOLANA_SECRET_KEY标为机密的值是你账户的钥匙。千万不要分享给别人,也不要粘贴到聊天里。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config,把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"agent-sniper": {
"command": "npx",
"args": [
"-y",
"@three-ws/agent-sniper",
"mcp"
],
"env": {
"SNIPER_MODE": "your-value",
"SNIPER_NETWORK": "your-value",
"SOLANA_RPC_URL": "your-value",
"SOLANA_SECRET_KEY": "your-value"
}
}
}
}在终端中运行这条命令
claude mcp add agent-sniper --env SNIPER_MODE=your-value --env SNIPER_NETWORK=your-value --env SOLANA_RPC_URL=your-value --env SOLANA_SECRET_KEY=your-value -- npx -y @three-ws/agent-sniper mcp运行前请先填写空着的值。
粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"agent-sniper": {
"command": "npx",
"args": [
"-y",
"@three-ws/agent-sniper",
"mcp"
],
"env": {
"SNIPER_MODE": "your-value",
"SNIPER_NETWORK": "your-value",
"SOLANA_RPC_URL": "your-value",
"SOLANA_SECRET_KEY": "your-value"
}
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"agent-sniper": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@three-ws/agent-sniper",
"mcp"
],
"env": {
"SNIPER_MODE": "your-value",
"SNIPER_NETWORK": "your-value",
"SOLANA_RPC_URL": "your-value",
"SOLANA_SECRET_KEY": "your-value"
}
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
SNIPER_MODE envSNIPER_NETWORK envSOLANA_RPC_URL env这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月6日
注册名称已通过 GitHub 账户 nirholas 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
你需要填写:SOLANA_SECRET_KEY
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(0.1.7)于37天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。
同一分类中有:Spec Driven Development, Remote Desktop Commander, yontu
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
Remote Desktop Commander: use files, terminal and processes on your own computers from any AI app.
Türkçe yazı denetimi: TDK kuralları, yapay zekâ izleri, metin türüne göre üslup.
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
A lightweight, embeddable pump.fun sniper engine for 3D AI agents. One trade loop — feed a stream of new launches, score each against a per-agent strategy, run the pre-trade guards, buy, then sweep open positions on an interval and exit on stop-loss / take-profit / trailing-stop / timeout / sentiment-flip. It is multi-agent and multi-user by construction (one wallet per agentId), and every external dependency — wallet custody, persistence, RPC/venue, launch feed — is a pluggable adapter, so the same engine runs locally with in-memory state or hosted across thousands of tenants. The package ships with four faces over that one engine: a library, a CLI, an MCP server, and an x402 paid HTTP API.
The engine is pure orchestration. It owns the decision pipeline; it owns nothing about *where* the money, the data, or the chain live. Five adapter interfaces (src/types.js) are the seams, and a default implementation ships for each:
┌──────────────────────────── engine ────────────────────────────┐
Feed ──{mint|intel|claim}─▶ score ──▶ guards ──▶ executeBuy ──▶ claimPosition │ (PumpPortal) (scorer) (concurrency, (quote + (atomic slot) │ budget, SOL build + sign │ headroom, + broadcast) │ price impact) │ position sweep (every pollMs) │ ──▶ decideExit ──▶ executeSell │ └────────────────────────────────────────────────────────────────────┘ adapters: Feed Solana Wallet Executor Store (launch (quote/build (resolve a (sign + (strategies, stream) pump.fun ix) Keypair per broadcast) positions, agentId) spend ledger)
Swap any one without touching the loop: a custodial KMS wallet instead of local keys, a Postgres store instead of memory, a different launch feed, a different venue. Implement the shape, pass it to createSniper(), done.
| Adapter | Responsibility | ||
|---|---|---|---|
Feed | Push `{ kind: 'mint'\ | 'intel'\ | 'claim', data: Candidate }` events into the engine. |
Solana | Quote + build pump.fun buy/sell instructions; expose a web3 Connection. | ||
Wallet | Resolve a signing Keypair for an agentId (or null if unprovisioned). | ||
Executor | The one place that signs and broadcasts; returns landing telemetry. | ||
Store | Persist strategies, positions, and the daily spend ledger; atomic claimPosition. |
Optional Hooks (firewall assessSafety, oracleGate, onScreen, onBuy, onSell, recordDecision) are best-effort cross-cutting callbacks — a throw is swallowed and never aborts a trade, except onTip, which is a deliberate veto.
npm i @three-ws/agent-sniper
Optional peers, installed only if you use the face that needs them:
npm i better-sqlite3 # durable Store (src/adapters/store/sqlite.js)
npm i express # the x402 HTTP API face
Requires Node >= 22.
The presets.local wiring is the zero-to-running path: in-memory store + self-custody wallet + pump.fun client + web3 executor + PumpPortal feed. Run it in simulate mode on devnet first — the engine scores, guards, and logs every decision but never broadcasts.
A strategy must carry a stop_loss_pct — the engine refuses to arm a strategy without one. SOL amounts on a strategy are expressed in lamports (1 SOL = 1,000,000,000 lamports), as a string, number, or bigint.
import { createSniper, createMemoryStore, presets } from '@three-ws/agent-sniper';
// 1 SOL = 1e9 lamports. const SOL = 1_000_000_000n;
const strategy = { id: 'strat_scout_1', agent_id: 'scout', // one wallet per agentId enabled: true, trigger: 'new_mint', // fire on the create event network: 'devnet', per_trade_lamports: (SOL / 100n).toString(), // 0.01 SOL per snipe daily_budget_lamports: (SOL / 2n).toString(), // 0.5 SOL/day ceiling max_concurrent_positions: 3, slippage_bps: 500, // 5% max_price_impact_pct: 10, // entry circuit breaker stop_loss_pct: 30, // REQUIRED — exit when down 30% take_profit_pct: 80, // exit when up 80% trailing_stop_pct: 20, // exit when down 20% from peak (after profit) max_hold_seconds: 1800, // hard time-stop require_socials: true, // skip launches with no twitter/telegram/website max_creator_launches: 10, // skip serial launchers };
const sniper = await presets.local({ network: 'devnet', mode: 'simulate', // no funds move strategies: [strategy], // secrets: { scout: ' ' }, // for live mode });
await sniper.start();
console.log(sniper.stats()); // { events, candidates, buys, sells, errors, strategies, queued, lastEventAt } console.log(sniper.strategies()); // the live armed-strategy cache
// Fire a manual buy by hand (drives the same path as MCP snipe_now): sniper.submitCandidate( { mint: 'THREEsynthetic1111111111111111111111111111111', entry_trigger: 'manual' }, { force: true }, );
// later… await sniper.stop();
createSniper(deps) is the lower-level constructor when you want to bring your own adapters — pass { config, store, wallet, solana, executor, feed, hooks }. presets.local is just one wiring of it.
The package installs an agent-sniper binary. Default mode is simulate — it will not move funds until you pass --mode live (or set SNIPER_MODE=live), which trades real funds.
# Run the engine. Mode defaults to simulate; live trades real funds.
npx agent-sniper run --network devnet --strategy ./strategy.json
agent-sniper mcp
agent-sniper serve
agent-sniper help
Safety.
run/serve/mcpall start in simulate mode unless you opt into
live explicitly. In live mode every armed strategy can spend up to its
daily_budget_lamportsper UTC day. Read Safety & guardrails below before
you flip the switch.
--strategy ./strategy.json loads a strategy (or array of strategies) as JSON. A minimal file mirrors the object in the Quickstart — note stop_loss_pct is required.
agent-sniper mcp exposes the engine as an MCP stdio server. Every tool maps 1:1 onto a public engine/store method — no internals are reached around. The server is unauthenticated and meant to run locally against your own wallet/RPC.
| Tool | Kind | Purpose |
|---|---|---|
arm_strategy | write | Register/update a strategy. SOL → lamports for you; stop_loss_pct mandatory. |
disarm_strategy | write | Disable a strategy (enabled=false); open positions still exit on their rules. |
list_strategies | read | List the armed strategy set the engine is evaluating. |
snipe_now | write | Force a manual buy of a mint across all armed agents (bypasses the scorer). |
list_positions | read | List positions, filter by agentId and/or status. |
close_position | write | Schedule an exit — flips the position's kill switch; the next sweep sells it. |
sniper_status | read | Engine health: event/candidate/buy/sell counts, armed count, queue depth, network/mode. |
Wire it into an MCP client (Claude Desktop, Cursor, etc.):
{
"mcpServers": { "agent-sniper": { "command": "npx", "args": ["-y", "@three-ws/agent-sniper", "mcp"], "env": { "SNIPER_NETWORK": "devnet", "SNIPER_MODE": "simulate" } } } }
createSniperMcpServer(deps) builds the fully-registered server without connecting a transport, so you can construct it in a test and drive it over an in-memory transport.
agent-sniper serve (or import { serve } from '@three-ws/agent-sniper/api') mounts the engine over HTTP. Reads are free; the four mutating endpoints are gated behind x402 USDC micropayments via @three-ws/x402-server — the middleware verifies the X-PAYMENT header against the facilitator, runs the work, settles on-chain, and emits the receipt.
| Method | Endpoint | Price | Notes |
|---|---|---|---|
| GET | /health | free | Liveness + network/mode/stats. |
| GET | /status | free | Full stats + immutable runtime config. |
| GET | /strategies | free | The armed strategy set. |
| GET | /positions?agentId=&status= | free | Positions from the store. |
| GET | /activity | free | The recent engine screen events (what the console feed renders). |
| GET | / and /console | free | The built-in operator console (see below). |
| POST | /strategies | $0.01 USDC | Arm a strategy (SOL → lamports; stop-loss required). |
| POST | /snipe | $0.05 USDC | Force a snipe on { mint, symbol?, agentId? }. |
| POST | /strategies/:id/disarm | $0.005 USDC | Disable a strategy. |
| POST | /positions/:id/close | $0.005 USDC | Schedule an exit; the next sweep sells it. |
Prices are the defaults (USDC atomic units, 6 decimals); override via deps.prices = { arm, snipe, disarm, close }.
Gating turns on when a merchant wallet is configured — deps.payTo.solana or the X402_PAY_TO_SOLANA env (a Base lane is available via X402_PAY_TO_BASE). When no payTo is set, the mutating routes still mount but answer 503, so the server boots cleanly for local/dev without payment config.
The Solana lane also needs a facilitator sponsor fee-payer — set X402_FEE_PAYER_SOLANA (or deps.feePayer). Without it the Solana lane self-disables (falling back to the Base lane, or the 503 guard if that's the only lane) rather than failing requests, and logs a one-time warning at boot.
import { serve } from '@three-ws/agent-sniper/api';
await serve( { payTo: { solana: process.env.X402_PAY_TO_SOLANA } }, { port: 8787 }, );
serve also ships a built-in operator dashboard — a single, dependency-free HTML file served from the same origin as the API, at / and /console. Open http://localhost:8787 and you get a live control room:
s snipe, n new strategy, r refresh, , settings, ? help), a disconnected banner, and full mobile responsiveness.Reads are free. The mutating controls (arm, snipe, disarm, close) are x402-paid for external callers — so to drive them from your own console without paying yourself, set an operator token:
export SNIPER_ADMIN_TOKEN="$(opens只显示了开头部分;其余内容请在仓库中阅读