About x402-tools-mcp
x402-tools-mcp is an MCP server with 22 pay-per-call tools for agent builders handling web content, data conversion, memory, and safety checks. Calls settle in USDC on Base through x402, without accounts or API keys.
Key features
- 01Web scraping and content conversion
- 02Embeddings and semantic memory retrieval
- 03Prompt-injection and URL risk checks
- 04JSON Schema and multi-agent chain validation
- 05Notifications, PDF creation, and moderation
Use cases
- 01Build a RAG flow from web scraping to vector storage
- 02Check retrieved content for prompt injection
- 03Validate API responses and URL safety
- 04Send workflow notifications or generate PDFs
MCP support
The MCP server exposes 22 tools for scraping, conversion, embeddings, storage, validation, notifications, and related tasks. Tools are paid per call in USDC on Base through x402.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @melis-ai/x402-tools-mcpX402_PAYMENT_HEADERSecretSigned x402 payment token from your wallet client. Set if handling payments manually.X402_INTERNAL_KEYSecretBypass key for trusted internal callers — skips x402 payment flow entirely.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
17Saved tool list · updated 3w ago
This tool takes no parameters.
Details
From the project README@melis-ai/x402-tools-mcp
Boring infrastructure for agent builders. 22 pay-per-call utility tools served as an MCP server — scrape, validate, embed, store, moderate, notify, convert, prevent loops. No accounts, no API keys, no subscriptions. Pay with USDC on Base via the x402 protocol.
npx @melis-ai/x402-tools-mcp
→ Full catalogue with prices and live examples: agents.melis.ai → Canonical 5-step RAG pipeline: agents.melis.ai/pipelines/rag → 5-minute getting started: agents.melis.ai/docs/getting-started
Tool catalogue (22)
| Tool | Price (USDC) | Description |
|---|---|---|
list_services | free | Returns the full catalogue with prices and endpoint URLs |
cacheserve | $0.001 | Fetch a URL with server-side caching — avoids redundant origin hits |
docconvert_text | $0.001 | Format conversion: md↔html, json↔csv, html→txt, etc. |
intentflow | $0.001 | Context handoff relay for multi-agent delegation (async, returns retrieve_url) |
memoryserve_write | $0.001 | Store text + embedding in Qdrant + SQLite, namespaced per agent_id |
memoryserve_query | $0.001 | Recall most-similar memories by semantic query |
memscrub | $0.001 | Scan retrieved RAG content for indirect prompt injection (10 patterns) |
notifyrelay_webhook | $0.001 | POST a JSON payload to a URL with optional HMAC-SHA256 signing |
schemagate | $0.001 | Validate a JSON response against a JSON Schema, return a hint on failure |
loopwall_hop | $0.0005 | Validate one hop of a signed multi-agent chain (loop / budget enforcement) |
loopwall_issue | $0.001 | Issue a signed Job Envelope at chain start with budget + hop cap |
imageguard | $0.002 | NSFW image classification (multi-class moderation on roadmap) |
notifyrelay_telegram | $0.002 | Send a Telegram message to a known chat ID |
promptguard | $0.002 | Score untrusted input for prompt injection risk (0–100) |
structextract | $0.002 | Extract structured JSON from HTML — tables, links, emails, phones |
xaudit | $0.002 | Validate API response content with a cryptographically signed certificate |
docconvert_pdf | $0.005 | Convert HTML or markdown to PDF (base64 output) |
kyaoracle | $0.005 | On-chain trust score (0–100) for an Ethereum/Base wallet address |
linkrisk | $0.005 | URL risk profile — heuristic phishing signals, redirect tracing |
markdownopt | $0.005 | Convert URL or HTML to clean LLM-ready markdown (~70% token reduction) |
notifyrelay_email | $0.005 | Send a transactional email via Resend (allowlisted, rate-limited) |
embedpay | $0.00005 / 1k tokens | OpenAI text-embedding-3-small wrapper, batch tier available |
linksafe | $0.01 | Definitive URL safety check via Playwright sandbox + VirusTotal |
scrapepay | $0.01 | Web extraction via Playwright (robots.txt enforced, SSRF-safe) |
pdf_render | $0.49 | High-fidelity URL or HTML → PDF via Playwright |
22 services. Live prices and live wallet on Basescan.
Why x402, why pay-per-call
- No signup tax. Every tool here works the moment your wallet has USDC. No accounts to provision, no API keys to rotate, no subscription tiers to forecast.
- Charge-on-success-only. Every service settles payment only on HTTP 2xx with non-empty content. Failed calls cost nothing — your retry budget is preserved.
- Composes naturally. ScrapePay → MarkdownOpt → EmbedPay → MemoryServe → MEMSCRUB is the canonical RAG pipeline. Each step's payment is independent and verifiable on-chain.
- MCP-native. One install, all 22 tools appear in Claude Desktop, Cursor, Cline, Continue, Claude Code, or any MCP-aware agent runtime.
- No vendor lock-in. x402 is a Linux Foundation protocol (April 2026) co-signed by Google, AWS, Stripe, Visa, Mastercard. Other providers will offer x402 services; your wallet client works with all of them.
Installation — Claude Desktop
Add to your claude_desktop_config.json:
{
"mcpServers": {
"x402-tools": {
"command": "npx",
"args": ["@melis-ai/x402-tools-mcp"]
}
}
}
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Restart Claude Desktop. All 22 tools appear in the tool picker.
Installation — Cursor, Cline, Continue, Claude Code
See agents.melis.ai/docs/install-mcp for per-client setup.
Example: canonical RAG pipeline
// All 22 tools registered as MCP tools — your agent calls them like any function
const page = await callTool('scrapepay', { url: 'https://example.com/article' });
const md = await callTool('markdownopt', { html: page.content });
const vec = await callTool('embedpay', { input: md.markdown });
const stored = await callTool('memoryserve_write', {
agent_id: 'agent-1', content: md.markdown, vector: vec.embedding,
});
// Later — recall and guard against indirect injection
const hits = await callTool('memoryserve_query', {
agent_id: 'agent-1', query: 'what did we learn about X?',
});
for (const chunk of hits.matches) {
const scrub = await callTool('memscrub', { content: chunk.content });
if (scrub.risk_level === 'safe') useChunk(chunk);
}
Total cost per 5k-token page: ~$0.017 USDC. Detailed walkthrough: agents.melis.ai/pipelines/rag.
Funding the wallet
x402 settles in USDC on Base (Coinbase L2). Two paths:
- Coinbase Wallet — buy USDC, switch to Base network, send to your agent's wallet.
- Bridge from Ethereum mainnet — bridge.base.org.
$1.00 of USDC covers 100 ScrapePay calls, 200 PromptGuard calls, or one full RAG-pipeline ingestion run.
Discovery surfaces (for agent builders + AI consumers)
| URL | Purpose |
|---|---|
| agents.melis.ai/llms.txt | Concise LLM-readable catalogue |
| agents.melis.ai/llms-full.txt | Full schemas + composition recipes |
| agents.melis.ai/openapi.json | OpenAPI 3.1 spec for all 22 endpoints |
| agents.melis.ai/ai-plugin.json | ChatGPT-plugin manifest |
https://<service>.melis.ai/.well-known/mcp.json | Per-service MCP discovery |
| agents.melis.ai/agents-faq | Agent-builder FAQ (distinct from human FAQ) |
Comparisons
- ScrapePay vs Firecrawl
- EmbedPay vs OpenAI Embeddings (direct)
- MemoryServe vs Pinecone
- Full alternatives doc
Source & licensing
- MCP wrapper source (this repo): MIT licensed.
- Service code (the 22 services themselves): closed source for security.
- Site content (agents.melis.ai docs): CC-BY-4.0 — attribute "melis x402 Tools" with a link to agents.melis.ai.
Security
Found a vulnerability? See SECURITY.md. Email [email protected]. Critical issues acknowledged within 24 hours.
Citing this work
If you use melis x402 Tools in research or downstream products, see CITATION.cff for the canonical citation.
Author
mizukaizen — agent operator and builder. Catalogue: agents.melis.ai · Email: [email protected]
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.