mcp.vanshul.com

A public Model Context Protocol (MCP) server, running as a Cloudflare Worker, that lets any AI agent read the live web as clean Markdown. It builds on the same extraction pipeline as the sibling reader/ project (Mozilla Readability + Turndown), exposed over the MCP Streamable HTTP transport so agentic clients can plug straight in.

Endpoint

POST https://mcp.vanshul.com/mcp     # JSON-RPC 2.0 (MCP)
GET  https://mcp.vanshul.com/health  # { ok: true, tools: [...] }

Tools

Tool	Input	Returns
fetch_markdown	{ url, max_chars? }	The page's main content as clean Markdown (optionally truncated to max_chars)
search_page	{ url, query, max_matches?, context_chars? }	Only the passages matching query, each with its heading breadcrumb, ranked by relevance — token-efficient alternative to fetch_markdown
fetch_metadata	{ url }	JSON: title, byline, siteName, excerpt, wordCount
extract_links	{ url, limit? }	JSON: all outbound http(s) links + anchor text

Connect from an MCP client

Remote/HTTP-capable clients (Claude Desktop, Cursor, Continue, …):

{
  "mcpServers": {
    "web-reader": { "url": "https://mcp.vanshul.com/mcp" }
  }
}

Stdio-only clients can bridge with mcp-remote:

npx mcp-remote https://mcp.vanshul.com/mcp

Add it to your client

• Cursor — Settings → MCP → Add new server, or drop this into ~/.cursor/mcp.json:

  { "mcpServers": { "web-reader": { "url": "https://mcp.vanshul.com/mcp" } } }
  

• Claude Desktop — add the same block to claude_desktop_config.json (Settings → Developer → Edit Config). If your version is stdio-only, use:

  { "mcpServers": { "web-reader": { "command": "npx", "args": ["mcp-remote", "https://mcp.vanshul.com/mcp"] } } }
  

• Continue / VS Code — add web-reader with URL https://mcp.vanshul.com/mcp to your MCP servers config.

Try it with curl

curl -s https://mcp.vanshul.com/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

curl -s https://mcp.vanshul.com/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"fetch_markdown","arguments":{"url":"https://example.com"}}}'

# Return only the passages matching a query (token-efficient):
curl -s https://mcp.vanshul.com/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call",
       "params":{"name":"search_page","arguments":{"url":"https://example.com","query":"more information"}}}'

Layout

mcp/
├── src/
│   ├── worker.ts     # entry: routes /mcp, /health, rate limit, CORS
│   ├── mcp.ts        # JSON-RPC dispatch + tool definitions
│   ├── extract.ts    # HTML -> Markdown / links (Readability + Turndown)
│   ├── search.ts     # query-focused passage search over extracted Markdown
│   ├── fetcher.ts    # fetch with timeout, size cap, re-validated redirects
│   └── security.ts   # SSRF guard (block private/internal addresses)
├── public/
│   ├── index.html    # landing page (served for non-API paths)
│   ├── og.png        # social share image (1200×630)
│   ├── og.svg        # social image source
│   ├── robots.txt
│   └── sitemap.xml
├── tests/            # vitest: security, extract, search, mcp dispatch, fetcher, worker
├── wrangler.toml
├── package.json
├── tsconfig.json
└── README.md

Develop & deploy

cd mcp
npm install
npm run typecheck
npm test          # vitest — security, extraction, search, MCP dispatch, fetcher, worker
npm run dev        # local worker at http://localhost:8787  (POST /mcp)
npm run deploy     # wrangler deploy

After the first deploy, attach the custom domain mcp.vanshul.com in the Cloudflare dashboard (Workers & Pages → this worker → Settings → Domains), or uncomment the [[routes]] block in wrangler.toml.

Security

• SSRF-safe: only public http(s) URLs; localhost, private ranges, cloud metadata and every redirect hop are blocked/re-validated.
• Bounded: 10s fetch timeout, ~3 MB page cap, max 5 redirects, per-IP rate limit.
• Stateless & private: no page content is stored; extraction is deterministic with no LLM in the loop.

MCP protocol

Implements initialize, ping, tools/list, tools/call and notifications (protocol version 2025-06-18). Tool failures are returned as { content, isError: true } so the agent can read the message and recover; malformed requests use standard JSON-RPC error codes.

Documentation

• docs/architecture.md — modules, request lifecycle, extraction pipeline, limits
• docs/tools.md — full tool & JSON-RPC API reference with examples
• docs/security.md — SSRF threat model and mitigations
• docs/deployment.md — Cloudflare Worker + custom-domain deploy guide

License

MIT © Vanshul Goyal