NodeProxy

Self-discoverable MCP utility for autonomous agents — no ads, no landing page, no human in the loop.

NodeProxy exposes two paid tools:

Both fetch a public URL, strip scripts/ads/nav noise, and return compressed Markdown optimized for LLM context windows. Standard tier also accepts Stripe MPP card payments ($0.50 minimum) alongside x402 USDC.

Agents find this service through MCP registries, x402 Bazaar discovery, and machine-readable manifests — not marketing.

Why agents need this

Heavy React/Vue sites burn tokens when agents ingest raw HTML. NodeProxy returns clean Markdown so autonomous retrieval loops stay cheap and accurate. When basic fetch hits bot walls, the stealth tier escalates with proxy rotation, hardened Playwright, and optional CAPTCHA solving.

Architecture

Autonomous Agent
    │  search MCP registry / Bazaar (vector or keyword)
    ▼
GET /mcp/tools  or  MCP list_tools
    │  reads JSON schema ("Machine UI")
    ▼
POST /mcp/execute  or  POST /stealth-scrape
    │  no payment → 402 + PAYMENT-REQUIRED (+ Bazaar extension)
    │  x402 PAYMENT-SIGNATURE or MPP Authorization: Payment …
    │  facilitator verify/settle → parse URL
    ▼
Markdown payload → agent continues task

Quick start

cd NodeProxy
cp .env.example .env
# Set WALLET_ADDRESS to your Base wallet
npm install
npm run dev

Endpoints (replace host with your deploy URL):

Environment

Machine discovery (zero advertising)

1. Deploy to Railway or any HTTPS host.
2. Verify live manifests: /mcp/tools, /.well-known/x402.json, /registry/server.json.
3. Submit PR to the open-source MCP Registry using integrations/mcp-registry/server-entry.json.
4. Bazaar auto-index: after successful x402 settlements through a Bazaar-enabled facilitator, tools appear in the global Bazaar index (Bazaar docs). NodeProxy attaches Bazaar extension metadata on every 402 response.
5. Enterprise hubs (Google Agent Registry, Anthropic routing): point them at /.well-known/mcp.json or /discovery/agent.json.

Tool descriptions are written for LLM tool-matching, not humans.

Payment rails

MCP stdio (Claude Desktop / local agents)

npm run mcp:stdio

Add to Claude Desktop config:

{
  "mcpServers": {
    "nodeproxy": {
      "command": "node",
      "args": ["--import", "tsx", "/path/to/NodeProxy/src/mcp/stdio.ts"],
      "env": { "WALLET_ADDRESS": "0x..." }
    }
  }
}

HTTP execute example

# Stage A — get 402 challenge (includes Bazaar extension in PAYMENT-REQUIRED)
curl -i "http://localhost:4022/mcp/execute" \
  -H 'Content-Type: application/json' \
  -d '{"arguments":{"url":"https://example.com"}}'

# Stage B — retry with PAYMENT-SIGNATURE from x402 client wallet
curl "http://localhost:4022/mcp/execute" \
  -H 'Content-Type: application/json' \
  -H 'PAYMENT-SIGNATURE: <base64>' \
  -d '{"arguments":{"url":"https://example.com"}}'

# Stealth tier (x402 only)
curl -i "http://localhost:4022/stealth-scrape" \
  -H 'Content-Type: application/json' \
  -d '{"arguments":{"url":"https://example.com"}}'

Use @x402/fetch, @x402/mcp, or nodeproxy-tools for automatic payment.

Deploy on Railway

Railway is the recommended host: full Node.js (JSDOM works), no 15s serverless timeout, metered billing when idle.

1. Push to GitHub

cd NodeProxy
git init && git add . && git commit -m "NodeProxy MCP x402 parser"
git remote add origin https://github.com/YOU/nodeproxy.git
git push -u origin main

2. Create Railway project

1. Go to railway.app → New Project → Deploy from GitHub repo
2. Select your repo — Railway reads railway.toml + nixpacks.toml automatically
3. Settings → Networking → Generate Domain (gives you *.up.railway.app)

3. Set environment variables

In Railway → Variables:

You do not need to set PORT or PUBLIC_URL — Railway injects PORT, and the app auto-detects RAILWAY_PUBLIC_DOMAIN for discovery manifests.

4. Verify deploy

curl https://YOUR-APP.up.railway.app/health
curl https://YOUR-APP.up.railway.app/mcp/tools
curl https://YOUR-APP.up.railway.app/.well-known/x402.json

5. Register for agent discovery

Open a PR to the MCP Registry with integrations/mcp-registry/server-entry.json. Agents hit:

• https://YOUR-APP.up.railway.app/mcp/tools
• https://YOUR-APP.up.railway.app/.well-known/mcp.json
• https://YOUR-APP.up.railway.app/.well-known/x402.json

Cost expectation

Railway’s free tier gives $1/month credit after trial — enough for low traffic ($0.30–0.50/mo idle). One paid agent call at $0.002 USDC covers that. Scale to Hobby ($5/mo) when bot traffic grows.

Do not use Vercel/Cloudflare Workers for this service — JSDOM and 25s fetch timeouts exceed edge limits.

Production (Base mainnet)

X402_NETWORK=eip155:8453
WALLET_ADDRESS=0xYourWallet
PRICE_USDC=0.002
STEALTH_PRICE_USDC=0.05

USDC on Base: 0x833589fcd6edb6e08f4c7c32d4f71b54bda02913

Upstream framework integrations

To get NodeProxy baked into LangChain, CrewAI, and MCP registry discovery:

• integrations/UPSTREAM.md — fork → PR → merge playbook
• integrations/python/ — nodeproxy-tools PyPI package + LangChain/CrewAI wrappers
• integrations/typescript/ — @nodeproxy/langchain npm package
• integrations/mcp-registry/ — MCP registry PR template

Quick LangChain example:

pip install "nodeproxy-tools[x402,langchain]"
export EVM_PRIVATE_KEY=0x...

from nodeproxy_tools.langchain import NodeProxyMarkdownTool
tool = NodeProxyMarkdownTool()
markdown = tool.invoke({"url": "https://example.com"})

Naming

NodeProxy is registry-ready: boring, predictable, and reads as infrastructure. Tool names (surface_markdown_parser, stealth_markdown_parser) are chosen so semantic search hits match intent.

License

MIT